Skip to content

Add eight-call cable routing with contact-aware fast pacing - #834

Merged
yuecideng merged 12 commits into
mainfrom
codex/cable-object
Nov 19, 2026
Merged

yuecideng merged 12 commits into
mainfrom
codex/cable-object

Conversation

@KeqiYe

@KeqiYe KeqiYe commented Oct 22, 2026 •

Copy link
Copy Markdown
Collaborator

Description

Run RouteTheCable-YAM-v1 through one canonical eight-call Task Program. The task routes the YAM-held cable around two spools, inserts the free end into the channel, releases it, and retracts both arms before final settling.

  • Replace the 57-stage / 148-invocation recipe with four contact-aware PickUp / Place pairs: 8 actual Atomic Skill invocations, 55 observed phase plans, 48 physical checkpoints and 45 internal holds. Continuation plans retain the invocation identity, timeout and row eligibility; they do not start nested skills.
  • Promote the combined optimization to the ordinary task.yam.yaml deployment. Remove the old recipe and temporary .compact.yaml deployment/program/integration/policy files. Keep the public task ID and standard CLI entry point.
  • Combine compact orchestration with phase-specific fast pacing. Physical task time decreases from 88–90 s to 44.60–45.77 s, approximately one half.
  • Retain four 0.35 s post-close dwells, 0.5 s soft-grasp settling, 0.25 s first lay-down settling, 1.0 s second lay-down settling, 0.5 s post-release settling and the final 2 s hold. Sensitive lay-down angular pacing stays at 0.5 rad/s. Ordinary motion uses 2.22× pacing to offset the additional contact settling.
  • Preserve the unified Newton/DexUni two-way solver, control/physics timesteps, VBD iterations, physical assets/materials and the existing lift, slip, contact, seating and strict endpoint criteria.
  • Propagate eligible masks before registered grounding. Unreachable IK, invalid lift tangents and wrong-way rays fail by physical environment ID with finite held fallbacks. Rejected rows no longer determine shared cable-move duration; healthy peers continue.

The PR also retains its managed USD cable import, YAM preset/assets and offline URDF override fixes, and retires the handwritten cable environment and separate GraspCable skill. The declared DexSim dependency and wheel locks remain aligned with main.

Clean task video — original speed (1×)

cable-task-clean-1x.mp4

Clean 1280×960 / 30 fps recording without captions, UI overlays or inset views. It shows environment 0 from the successful final N=2 run: 44.60 s of physical task execution + 5 s of additional live-physics persistence, approximately 49.6 s total at 1× playback. Images are rendered from the recorded native robot/cable states; rendering adds no simulation steps. This video uses the final one-second second-lay-down settle.

Final physical qualification

Fresh processes, seed 73, RTX 5090 and DexSim 0.6.0rc2 build dbaca660471d5aaa7e750093e7fe1c207465c4af. N=2 and the first N=16 use 1c3080fac; the independent N=16 repeat uses final head cbbee29a2, whose only subsequent change is task README documentation.

Case Result Actual calls Control steps Physical task time Maximum slip
N=2 2/2 8 2676 44.60 s 0.861 mm
N=16 16/16 8 2746 45.77 s 1.387 mm
N=16, fresh repeat 16/16 8 2745 45.75 s 2.849 mm

34/34 completed all eight calls. Every released endpoint stayed inside the unchanged bounds [0.5845, 0.064, 0.753] to [0.5875, 0.0896, 0.759] m through release, retreat, home, settling and five additional physical seconds, with zero outside samples in every row. Task times include the final two-second hold and exclude those extra five seconds. All 48 physical checkpoints and 55 phase-completion gates passed in each environment.

The normal-pacing eight-call baseline took 88.20 s for N=2 and 90.10 s for N=16. Actual invocation count decreases from 148 to 8 (94.6%); physical duration decreases by approximately 49–50%. No playback acceleration is used to obtain those timings.

The unchanged 12 mm slip and 1 mm penetration limits pass. Final sampled cable/bench, cable/spool and cable/port penetration is zero. No cross-arena contact rows or dropped query rows were observed. A selected-row reset leaves other cable states, robot positions/velocities and targets unchanged and advances no physics time.

These fixed-seed batches qualify the tested setup; they are not a multi-seed success-rate estimate. An earlier main-integrated batch at a5cd502ee was 15/16 because one endpoint transiently rose above the strict release bound during the second lay-down. The release gate correctly rejected it. The recorded failure and diagnosis remain visible. Extending only that contact settle to one second fixes the observed failure in the fresh batches above without loosening acceptance bounds. Earlier short-settle results are superseded by this table.

Integration and validation

Merged main 22946b0d3 into this branch, retaining main's phase-path support for ordinary rigid PickUp and its inactive-field fingerprint behavior. Observed contact paths explicitly reject incompatible phase-path overrides.

  • 1141 CPU tests passed for affected Atomic Skills, configured integration, Gym bridge, metadata, task catalog/package resources and agent context.
  • 776 CPU tests passed for Task Program compilation/runtime and cable/contact behavior; GPU-marked cases were selected separately.
  • 123 focused CPU tests passed after the final settling change, including vertical-slice fingerprints, cable material geometry and observed contact paths. The fingerprint update was checked causally against the intentional public goal-descriptor expansion.
  • 3/3 fresh-process CUDA regressions passed at the main-integrated implementation: first-spool release, third carry and the complete route with opening/retreat/home, independent channel bounds, final settling and five further physics seconds. The complete single-environment case took 43.95 physical seconds before the final settling adjustment; the final profile is qualified by the full N=2/N=16 batches above.
  • Canonical deployment preflight: 8 segments / 8 calls.
  • Final local wheel build passes (pip wheel --no-deps --no-build-isolation .); package resource coverage passes.
  • Black 26.10.1, public API documentation coverage (5187/5187), context and whitespace checks pass. Sphinx dummy build succeeds with existing repository warnings.
  • Focused coverage protects per-row unreachable geometry, non-contiguous physical IDs, inactive-row duration independence, observed continuation/rejection, no rigid cable attachment, critical waits and the original physical acceptance limits. Legacy check fixtures are isolated test programs rather than a retained runnable task recipe.

The task README documents the canonical entry point and pacing; affected Atomic/Task Program context reflects the actual ownership and eligibility boundary.

Remote lint, documentation tests and automated review pass on cbbee29a2. The full-repository test/build run 44558053118 is still unfinished at publication; this is not an all-green full-CI claim. The repository has no required status checks and automatically cancels unfinished PR runs after merge. Merge validation is the focused CPU/CUDA, complete native parallel-task acceptance and local wheel build reported above.

Type of change

  • Enhancement / new feature
  • Bug fix
  • Breaking change: retire the handwritten environment and duplicate grasp skill
  • Documentation update

Checklist

  • Run black . with Black 26.10.1.
  • Review and update affected documentation and project context.
  • Document public API changes and pass the API coverage gate.
  • Review test necessity and validate focused behavior and full native physics.
  • Keep dependency/wheel declarations aligned with main.

@KeqiYe KeqiYe added enhancement Improvements to existing capabilities; use feature for new capabilities. physics Physics backends, dynamics, contacts, fluids, deformables and physics solvers. task Concrete tasks, deployments, objectives and recipes, including configured Task Program tasks. assets Robot and scene assets, meshes, materials, asset catalogs and downloads. labels Oct 22, 2026
@yuecideng
yuecideng self-requested a review October 22, 2026 06:37
@KeqiYe
KeqiYe force-pushed the codex/cable-object branch from 4e10a42 to f890c7e Compare October 22, 2026 08:26
@KeqiYe KeqiYe changed the title Add cable simulation support and routing demos Add cable grasp skill and YAM routing demos Oct 22, 2026
@KeqiYe KeqiYe added robot Robot runtime, presets, control groups, calibration and robot support. atomic-skill Atomic Skill goals, planning, execution, verification and recovery. labels Oct 22, 2026
@KeqiYe
KeqiYe changed the base branch from main to codex/cable-framework October 22, 2026 08:29
@yuecideng yuecideng added feature New user-facing capabilities. bug Incorrect existing behavior or a fix restoring the intended behavior. compute Shared numerical algorithms and kernels independent of simulation objects. docs Documentation, API references and example explanations. kinematics FK, IK, Jacobians, reachability and kinematic solvers; not physics solvers. and removed enhancement Improvements to existing capabilities; use feature for new capabilities. labels Oct 25, 2026
@KeqiYe
KeqiYe force-pushed the codex/cable-object branch from d97de17 to da1dabd Compare November 9, 2026 02:20
@KeqiYe
KeqiYe force-pushed the codex/cable-framework branch from 485dfd5 to 6ca6566 Compare November 9, 2026 02:20
@KeqiYe KeqiYe changed the title Add cable grasp skill and YAM routing demos Add cable grasping and YAM routing on DexSim 0.6 Nov 9, 2026
@KeqiYe
KeqiYe marked this pull request as ready for review November 9, 2026 05:36
@greptile-apps

greptile-apps Bot commented Nov 9, 2026 •

Copy link
Copy Markdown

RetriggerView in GreptileConfidence Score: 5/5

[High impact] The README-only update appears safe to merge.

Summary

The PR adds an eight-call cable-routing program with contact checks and faster motion. This follow-up only updates its README.

  • Pickups and placements continue through observed contact phases within one skill call.
  • The cable route uses measured material points and per-row checks to guide its eight-call program.
  • USD scene imports can now include deformable cables.

Reviews (4) · Last reviewed commit: "docs: document canonical cable pacing an..." · Reviewed by Greptile

@yuecideng
yuecideng requested a review from han-xudong November 9, 2026 05:42
@yuecideng yuecideng added motion gen Robot motion planning, MotionGenerator, trajectory generation, retiming and expansion. task-program Task Program language, semantics, compiler, runtime and integration contracts. labels Nov 13, 2026
@yuecideng

Copy link
Copy Markdown
Contributor

Task Program 迁移已更新;完整入槽仍失败,请原作者继续修复

@KeqiYe,本次迁移提交为 1a8bc308d905c174a25da007045a48723038ead9。请接手剩余入槽与释放稳定性的修复,当前仍不能将该任务标记为完整成功。

代码变化:删除 route_the_cable.py 自定义环境类,由配置注册通用 EmbodiedEnv;49 个阶段 / 125 次调用统一走 Task Program,抓取继续复用 PickUp(ContactGraspGoal),运输/开合复用已有 MoveEndEffector / MoveJoints。闭爪最后一帧被 env.step() 消费后才记录材料参考,运输阶段持续检查滑移,最终验收使用命名物理规则。求解器保持 DexUni 统一链路。

本次实现前后对比

下面是本次迁移的独立成对复测,基线为 694ddda7d,迁移版为 1a8bc308d;不是上一轮 GraspCable → PickUp 对比的数据。

指标 迁移前:手写环境 迁移后:Task Program
最大材料滑移 12.0343 mm 7.7618 mm
停止位置 lower_channel_0,滑移超过 12 mm lower_channel_3_tracking,TCP 跟踪超差
控制步数 3,271 5,282
仿真时间 54.52 s 88.03 s
strict_port False False
完整 task_success False False

这是单种子各一次运行。滑移更小或推进到更后的阶段,均不能据此推断成功率提升或根因已经确定。新执行策略统一应用关节加速度预算,动作时间更长;场景/机器人物理参数一致。episode 上限由 6,000 提高到 12,000 步,两次失败均未触发该上限。

当前明确失败点

迁移版完成三次绕柱运输、左手抓取和前几轮对槽修正后,在最后一层下降的实测检查处停止:

stage: lower_channel_3
call: simulation.cable_check
condition: lower_channel_3_tracking
position_error_m: 0.0072465353  (limit: 0.006)
orientation_error_rad: 0.0156860761
max_grasp_slip_m: 0.0077618398  (limit: 0.012)
measured tip XYZ, m: [0.58577466, 0.07640365, 0.76911253]
strict_port: False
release_ready: False
task_success: False

该端点的 Z 比严格槽内上界 0.759 m 高约 10.11 mm。此时尚未执行 release_free_end / settle_and_validate;不能用 broad end_in_port=True 替代严格入槽成功。

其他诊断:两柱 seat count 为 4 / 3,port count 为 9;线缆最大速度约 0.00910 m/s;bench/spool/port 穿透约 0.06694 / 0.00798 / 0 mm;机器人/绕柱接触查询未报告穿透。当前停止的直接触发项是 TCP 平移跟踪,不能仅靠通过滑移/穿透检查判定成功。

此前的释放时脱槽报告仍需处理。本次样本在释放之前已停止,因此它补充了当前失败路径,不能证明先前的释放后脱槽已经修复。

请继续修复的范围与验收

更新前 GitHub 已显示本 PR 与当前 main 有合并冲突。本次只追加迁移提交,没有额外合入新 main;后续还需处理该既有集成状态。

  1. 对照 lower_channel_0..3 的命令 TCP、实测 TCP、材料点和远端轨迹,定位最终下降时的持续位置偏差。检查接触载荷、夹爪/槽口几何和目标补偿;同时核对旧专家逐帧停止、附加跟踪等待与新 Task Program 有限修正/保持时序的差异。请先用轨迹/接触证据定位,暂不把原因归结为单一参数。
  2. 在 6 mm TCP、12 mm 滑移、1 mm 穿透及原 strict-port 约束下,使 seed 73 完整通过对槽、入槽、开爪、撤离及 2 s 稳定,最终 task_success=True。保留 DexUni 统一求解器;不要以放宽容差、移除检查或只检查释放前位置来替代修复。
  3. 保留首柱释放及第三次运输的既有 NoRender CUDA 回归,并增加/完善完整路线回归,记录开爪后、撤离后和稳定后的实际端点。后续再补多 seed、多 episode 的成功率与失败分类。

当前代码定位:

复现设置及已验证范围

  • 一环境、seed 73、60 Hz 控制;使用 task.yam.yaml,遍历通用 env.create_demo_segments(),逐帧 env.step(action),按段执行 validator。
  • 本次 FastRT 录像:960×720,intrinsics (830, 830, 480, 360),eye (1.5, -1.5, 1.65),target (0.42, -0.05, 0.86);每四个控制步采图,渲染前显式 sim.sync_render_state(),采图不额外推进物理。默认任务配置为 NoRender,录像时显式切换 FastRT。
  • DexSim a3daf24edd88b489702c32ccffc7ffa405741ab2,Python 3.12、Torch 2.11.0+cu130、Newton 1.7.0、Warp 1.17.0,RTX 5090。不能假定任意未资格验证的 wheel 都能复现相同轨迹。
  • 本次物理参数摘要 SHA-256:b18c80ab3711041f9e59ebea92ca55b16bde5aac56a0217f6ea3d26f311fbb31,对比的是环境物理配置及机器人 simulation 配置,排除 episode 步数预算;与上一轮报告的摘要口径不同,不直接横向比较哈希。
  • 2,468 项相关 CPU 回归通过;生命周期 hook 提取后又复核 204 项。两项 NoRender CUDA 用例通过:首柱释放滑移 10.33 mm,第三次运输滑移 4.76 mm,后者机器人/柱体最小接触距离 0 mm。这些通过项仅资格验证各自中间阶段,不代表完整任务通过。
  • Black 26.10.1、部署 inspector(49 段 / 125 调用)、API 覆盖 4,880/4,880、context 检查通过。本次推送前重新检查静态门禁;上述运行结果对应此次迁移代码。

2× 同步对比视频、逐帧命令/实测机器人及线缆状态、阶段结果和源码快照已保留在本次 Codex 对话的对比预览中。PR 内的表格、失败日志和提交链接可独立用于继续定位;完整入槽验收目前仍为失败。

@yuecideng yuecideng left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Re-reviewed 1955c6402...2e502cf79 after the author's changes. Four confirmed P2 findings remain in the inline comments. Please address them before merge; this is a COMMENT review, not an approval or a REQUEST_CHANGES event.

Priority Location Defect Impact Recommended action
P2 move_joints.py:180 Authored samples bypass speed qualification and observed-start transition 60 rad/s accepted against a 1 rad/s cap Preserve cadence while qualifying all transitions
P2 pick_up.py:348 Shared declaration changes leave drawer golden expectations stale Required fast CI assertion fails Reconcile independently verified fingerprint expectations
P2 simulation/cable.py:606 Full-batch condition includes already rejected rows One failed cable aborts healthy environments Use row-local failure/eligibility propagation
P2 cable_two_posts.py:186 Both new demos omit visualization settings Required example-coverage CI assertion fails Supply visualization settings and launcher plumbing

The previously reported release-induced endpoint escape did not reproduce in two fresh complete runs of this head. I kept the configured strict bounds, 12 mm slip limit, 1 mm penetration limits and the unified DexUni solver path unchanged.

Physical verification Result
Runtime DexSim 0.6.0rc2, build dbaca660471d5aaa7e750093e7fe1c207465c4af, CUDA, one environment, seed 73
Original full-route CUDA test, NoRender PASS; natural generator exhaustion, release, retreat and 2 s settling; 333.75 s wall time
Independent full-route FastRT run All 51 segments PASS, 5,987 control steps, task success true; approximately 99.78 s simulated task time
Strict endpoint bounds (m) [0.5845, 0.064, 0.753] <= endpoint <= [0.5875, 0.0896, 0.759]
NoRender final endpoint (m) [0.584582865, 0.076319724, 0.753657103]
FastRT final endpoint after extra 5 s (m) [0.584591210, 0.077162221, 0.753960550]
Release through retreat, full opening, home and extra settling 2,118 live samples; zero samples outside strict channel bounds
Extra 5 s endpoint coordinate ranges X 0.0012 mm; Y 0.0093 mm; Z 0.0231 mm
Maximum grasp slip NoRender 0.353 mm; FastRT 0.299 mm, both below 12 mm
Reported penetration Robot/obstacle solver checks pass; final cable bench/spool/port metrics each 0 m

The extended check advanced physics with the last robot command held and read fresh cable metrics every physics step; it did not use sealed task-success state as a substitute for physical persistence. Captures synchronized renderer state and did not advance physics. This verifies the fixed deployment at seed 73, not randomized or multi-environment physical robustness. The installed SDK is newer than the author's qualified 0.6.0.dev12034 build; that build was not rerun here.

Review item Result
Target PR #834, 1955c6402...2e502cf79
Scope 54 changed files; Atomic Skills, configured Task Program/runtime, material geometry/evaluation, deployment and tests; source Sphinx pages were not opened under repository instructions
Execution Multi-agent independent static/CPU passes plus lead reproductions and native CUDA/FastRT execution
Findings P0: 0; P1: 0; P2: 4; P3: 0
Merge assessment Changes requested (P2 findings); remote event remains COMMENT
Validation Configured integration/bridge/cable suites 348 passed; selected Atomic cases 18 passed; verification slice 71 passed / 2 failed; cylinder checks 5 passed; deployment preflight clean; full-route native runs pass
Residual risk Physical execution limited to one environment, seed 73 and installed rc2 SDK; randomized configurations and other native SDK builds not qualified
Role Scope Executor / mode Status / evidence Candidate disposition
architecture Task Program/API/Atomic contracts and row continuation Independent architecture agent Complete static/CPU; 348 + 18 selected passing cases; engine speed probe Two accepted; row finding merged with robotics
robotics Geometry, frames, contact/release and material tracking Independent robotics agent Complete static/CPU; 55 selected passing cases; mixed-row probe One accepted and merged with architecture
verification Evaluation truth, regression gates and acceptance-oracle quality Independent verification agent Complete static/CPU; 71 pass / 2 fail, base fingerprint causality verified Two accepted

The lead independently reproduced all four findings, the base/head drawer difference and both existing CI assertion failures. The CI fast phase records 14,583 passes, 58 skips and these 2 failures in job 131898328579; its native phase was not reached. Broader AI/training and security specialist passes were not selected because this delta does not introduce those execution surfaces. Context additions match the inspected code, so no context correction is requested.

Comment thread examples/sim/demo/cable_two_posts.py Resolved
@yuecideng

Copy link
Copy Markdown
Contributor

Parallel cable routing: reproduction, confirmed root causes, and proposed correction

@KeqiYe, the completed native diagnostics identify two P2 issues: endpoint correction saturates at the TCP height floor, and row-local measured failures propagate as a batch-wide preparation error. Please address both while preserving the unified DexUni solver and the current physical acceptance criteria.

Revision and evidence scope

All measured runs below used 367fc0381f3e84b8116f2c6a903ad4b88b225af5 with the original initial posture. The current PR head is 0cee7b61ca59b43ce2113f3ae5a4062b9316e2ef; its startup-posture adjustment has not been rerun here. The task's integration.yaml and simulation/cable.py are byte-for-byte unchanged between these revisions. The numbers below are therefore evidence from the validated revision, not a claim that the new posture has already failed or passed.

Runtime: RTX 5090, Python 3.12, CUDA, NoRender, seed 73, DexSim 0.6.0rc2, build dbaca660471d5aaa7e750093e7fe1c207465c4af. Each normal batch was a fresh process driving the configured YAM task through production execute_demo_episode(). Diagnostic observations did not advance physics or write physical state.

Run Observed result
Normal 2 environments 2/2 completed all 51 stages, 5,974 control steps. After the configured 2 s settling interval, another 5 physical seconds were advanced holding the last command. Each row had 2,081 release-through-extra-settling endpoint samples; none left the strict region. Maximum grasp slip: 0.362 mm.
Normal 16 environments All rows passed the first 45 stages. At control step 5,228, verify_release_alignment failed for rows 4/15 and the complete batch was marked failed, including the 14 release-ready rows. Opening/retreat/settling were not reached.
Independent 16-environment diagnostic repeat Same failure stage at step 5,194; this time rows 7/8/12 were unseated and the other 13 rows were release-ready. Captured target-builder inputs/outputs demonstrate Z saturation.
Separate 2-row observer-fault injection Setting only row 0's cumulative-slip observation to 13 mm rejected that row, but a later check also aborted healthy row 1. This synthetic case is excluded from normal physical success counts.

Environment IDs are zero-based. The changing failed row IDs are consistent with run-dependent contact/tracking/deformation states; the repeat reproduces the control/propagation mechanism, not identical trajectories. These are not multi-seed success-rate measurements.

1. Endpoint feedback loses downward authority at the terminal TCP floor

The new inline finding anchors the configuration. Relevant implementation: target correction and clamping.

lower_channel_3 sets commanded TCP Z to 0.7655 m. The last two tip_align targets request endpoint Z = 0.754 m, but retain minimum_z: 0.7655. The calculation is:

requested_dz = clamp(0.754 - measured_endpoint_z, -0.003, +0.003)
next_tcp_z  = max(commanded_tcp_z + requested_dz, 0.7655)

Once the commanded TCP is already at the floor, a nonzero downward endpoint correction is reduced to a no-op. This is observed before the command reaches the robot, not inferred only from a final task failure.

Last-alignment measurements from the diagnostic repeat:

Row Requested Z delta (mm) Effective commanded Z delta (mm) Final endpoint Z (m) TCP position error (mm)
7 -3.000 +0.000119 0.759316 0.025
8 -3.000 -0.001490 0.760260 2.370
12 -3.000 -0.001550 0.759842 2.507

The effective deltas are essentially zero; their micrometre-scale differences reflect command/FK/IK residuals. In the earlier run, row 4's endpoint was [0.586565197, 0.076845646, 0.759543240] m and row 15's was [0.586382866, 0.076781750, 0.761865616] m: respectively 0.543 and 2.866 mm above the strict upper bound.

The controlled tool and the flexible cable endpoint are different quantities. The grasp tracks material coordinate 0.57 m, leaving approximately 30 mm of distal cable. In row 7, TCP tracking was accurate to 0.025 mm, but that free span placed the endpoint approximately 0.797 mm above the grasped material point, exceeding the channel bound. Rows 8/12 additionally had approximately 2.1 mm of vertical TCP tracking offset under load. Their TCP position/orientation checks still passed the configured 6 mm / 0.25 rad tolerances.

The final seated guard correctly blocks release. The missing behavior is earlier: after a fixed number of saturated correction slots, the controller has neither restored useful endpoint control nor produced a row-local recovery/infeasibility result. More identical correction slots alone cannot restore that control authority.

2. A physical condition failure is represented as a global grounding exception

This extends the existing row-isolation finding and its native parallel confirmation, rather than adding a duplicate finding.

_CableCheckLowerer.lower() reduces a per-row passed vector using passed.all() and raises one ValueError. SemanticCallExecutor._fail_preparation() then applies:

failed = self._eligible.clone()
self._failed |= failed
self._eligible &= ~failed

Consequently, two unseated rows aborted 14 ready rows in the first 16-row run; three aborted 13 in the diagnostic repeat. Registered lowerer invocation also lacks the executor's eligibility mask, so already-rejected rows can cause the next full-service-batch check to throw again.

The separate native fault-injection case makes the latter path explicit: after row 0 was rejected for the injected 13 mm observation, healthy row 1 had only 0.035 mm real material slip, 0.836 mm TCP error and 0.0032 rad orientation error. Nevertheless, raise_0_tracking threw an exception naming env 0 while the preparation-failure mask was [False, True], aborting the remaining healthy row.

Copy-paste reproduction

Use the qualified YAM asset from this PR and an environment with the pinned rc2 SDK. Pin the previously tested revision to reproduce the measurements; rerun the same cases on the latest initial posture when validating the correction.

git fetch origin refs/pull/834/head
git worktree add --detach /tmp/pr834-repro-367fc03 367fc0381f3e84b8116f2c6a903ad4b88b225af5
cd /tmp/pr834-repro-367fc03
export PYTHONPATH="$PWD:$PWD/extensions:$PWD/extensions/embodichain_tasks:$PWD/extensions/embodichain_models:$PWD/extensions/embodichain_benchmarks:${PYTHONPATH:-}"

# Save the Python block below as /tmp/pr834_parallel_repro.py.
# Normal physical batches, each in its own process:
python /tmp/pr834_parallel_repro.py --num-envs 2 --output /tmp/pr834-env2.json
python /tmp/pr834_parallel_repro.py --num-envs 16 --output /tmp/pr834-env16.json

# Read-only target-builder instrumentation:
python /tmp/pr834_parallel_repro.py --num-envs 16 --trace-tip --output /tmp/pr834-tip-trace.json

# Separate synthetic observer-fault case; does not modify physical state:
python /tmp/pr834_parallel_repro.py --num-envs 2 --fault-row 0 --output /tmp/pr834-fault.json

Look for release_alignment failures, the per-row preparation-failure mask in the saved segment runtime metadata, and TIP_TARGET rows with seated=False, a negative requested_dz, and essentially zero effective_dz. Do not require the exact same failed environment IDs.

Minimal reproduction script (same config/reset/lazy production-executor path; optional read-only target trace)
from __future__ import annotations

import argparse
import json
import subprocess
import dexsim
from pathlib import Path

import gymnasium as gym
import torch
import yaml
import embodichain_tasks
from embodichain.lab.gym.envs.demo import execute_demo_episode
from embodichain.lab.gym.utils.gym_utils import config_to_cfg
from embodichain.lab.sim import SimulationManager
from embodichain.lab.task_program.integrations.simulation.cable import _CableRuntime


def main() -> int:
    parser = argparse.ArgumentParser()
    parser.add_argument('--num-envs', type=int, default=16)
    parser.add_argument('--seed', type=int, default=73)
    parser.add_argument('--trace-tip', action='store_true')
    parser.add_argument('--fault-row', type=int, default=-1)
    parser.add_argument('--output', type=Path, default=Path('route834-repro.json'))
    args = parser.parse_args()
    if not -1 <= args.fault_row < args.num_envs:
        parser.error('--fault-row must select an existing environment, or -1')
    path = Path('extensions/embodichain_tasks/configs/tasks/manipulation/cable/route_the_cable/task.yam.yaml').resolve()
    cfg = config_to_cfg(yaml.safe_load(path.read_text()), source_path=path)
    cfg.num_envs = args.num_envs
    cfg.seed = args.seed
    cfg.sim_cfg.headless = True
    cfg.sim_cfg.render_cfg.renderer = 'no-render'
    cfg.sim_cfg.robot_ik_gizmo = None
    cfg.sim_cfg.enable_entity_gizmo = False
    report = {'head': subprocess.check_output(['git', 'rev-parse', 'HEAD'], text=True).strip(), 'sdk_version': dexsim.__version__, 'sdk_build': dexsim.__commit_id__, 'num_envs': args.num_envs, 'seed': args.seed, 'stages': [], 'tip_targets': []}
    env = None
    original_pose = _CableRuntime.pose

    def trace_pose(runtime, target, part):
        if target.kind != 'tip_align' or target.position[2] is None:
            return original_pose(runtime, target, part)
        before = runtime.tool(part, commanded=True)
        tip = runtime.metrics()['end_position']
        seated = runtime.seated()
        goal = original_pose(runtime, target, part)
        row = {
            'target': target.target_id,
            'tip_z': tip[:, 2].cpu().tolist(),
            'seated': seated.cpu().tolist(),
            'command_tcp_z': before[:, 2, 3].cpu().tolist(),
            'goal_tcp_z': goal[:, 2, 3].cpu().tolist(),
            'requested_dz': (target.position[2] - tip[:, 2]).clamp(-target.maximum_z_correction, target.maximum_z_correction).cpu().tolist(),
            'effective_dz': (goal[:, 2, 3] - before[:, 2, 3]).cpu().tolist(),
        }
        report['tip_targets'].append(row)
        print('TIP_TARGET', json.dumps(row), flush=True)
        return goal

    def segments():
        yield from env.create_demo_segments()

    def on_segment(result):
        runtime = env._active_task_program_bridge.post_policy_port._cable_runtime
        metrics = runtime.metrics()
        row = result.to_metadata()
        row['tip'] = metrics['end_position'].cpu().tolist()
        row['release_ready'] = metrics['release_ready'].cpu().tolist()
        report['stages'].append(row)
        print('SEGMENT', result.name, list(result.successes), flush=True)
        if args.fault_row >= 0 and result.name == 'right_pinch_0.162':
            # Synthetic observer fault ONLY; do not write physical state.
            runtime.max_slip[args.fault_row] = 0.013
            report['synthetic_fault'] = {'row': args.fault_row, 'cumulative_slip_m': 0.013}

    try:
        if args.trace_tip:
            _CableRuntime.pose = trace_pose
        env = gym.make('RouteTheCable-YAM-v1', cfg=cfg).unwrapped
        env.reset(seed=args.seed)
        result = execute_demo_episode(env, segments=segments(), on_segment_completed=on_segment)
        report['episode'] = result.to_metadata()
        runtime = env._active_task_program_bridge.post_policy_port._cable_runtime
        report['final_metrics'] = {key: value.cpu().tolist() for key, value in runtime.metrics().items()}
        print('SUCCESS_BY_ENV', list(result.success), flush=True)
        print('FINAL_METRICS', json.dumps(report['final_metrics']), flush=True)
        return 0 if result.all_success else 1
    finally:
        _CableRuntime.pose = original_pose
        args.output.parent.mkdir(parents=True, exist_ok=True)
        args.output.write_text(json.dumps(report, indent=2))
        if env is not None:
            env.close(exit_process=False)
        env = None
        SimulationManager.flush_cleanup_queue()


if __name__ == '__main__':
    raise SystemExit(main())

The full diagnostic runs also queried robot/fixture contacts and recorded complete per-stage cable/robot states. The compact script above uses the same execution entry points; it has been syntax/import/CLI checked, without launching another full batch for this publication. The normal 2-row extended-settling result came from the full diagnostic harness, not an extra-settling implementation in this compact script.

Proposed correction

  1. Make terminal endpoint control constraint-aware. Detect a substantial endpoint error whose requested correction is saturated away. Choose a bounded motion/recovery that can actually reduce the measured endpoint error while respecting TCP-height and collision constraints, or return an explicit row-local infeasible outcome. Qualify the geometry/contact consequences of any altered approach, grasp or wrist pose. Simply lowering the floor has not been established as a safe solution by these experiments.
  2. Use endpoint convergence to finish terminal alignment. Keep TCP tracking as a separate condition. Re-evaluate actual endpoint position after bounded corrections; an accurate TCP alone does not establish flexible-tip insertion.
  3. Preserve row-local failure semantics. Carry eligibility through the relevant registered-check boundary and express measured physical rejection as typed per-row qualification/outcomes. Healthy rows should continue. Genuine configuration/grounding errors can retain global preparation-failure handling.
  4. Keep the existing DexUni path and physical criteria. No alternate solver, runtime attachment, physical-state teleport or relaxed acceptance threshold is proposed.

Acceptance checks

  • Reproduce an unseated endpoint at the terminal TCP floor. Verify that nonzero endpoint error leads to feasible recovery or an explicit row-local failure, rather than exhausting nominally successful no-op alignment slots.
  • On the latest startup posture, complete normal 2-row and 16-row routes through full opening, retreat, home and physical settling. Require every successful row to satisfy the unchanged strict region [0.5845, 0.064, 0.753]..[0.5875, 0.0896, 0.759] m, grasp slip <= 12 mm and penetration <= 1 mm. Retain the configured 2 s settling and add a 5 s live-physics persistence check using fresh endpoint measurements.
  • Exercise mixed seated/unseated rows and a rejected/healthy-row case. Only the measured failing rows should fail; healthy rows must be able to complete the remaining task. Verify that rejected rows do not poison subsequent checks.
  • Preserve partial-reset isolation. The existing 16-row diagnostic already verified that resetting only row 0 left all other cable states and robot qpos/qvel/targets bitwise unchanged, without advancing the physics clock.
  • Repeat fresh-process qualification and report per-row successful/attempted counts. These diagnostics do not isolate a particular DexUni kernel defect or establish broad randomized robustness.

@yuecideng

Copy link
Copy Markdown
Contributor

Independent acceptance at d2a182945: normal routing passes; survivor continuation remains incomplete

@KeqiYe, I reran the updated deployment on the same rc2 SDK. The normal 2/16-environment routing and release cases pass. The complete rejected/healthy-row case still fails, so I cannot mark the row-isolation issue fully resolved.

Confirmed native results

Fresh-process case Result Evidence
Normal N=2 2/2 PASS All 57 stages, 5,302 controls. Each row had 2,001 release-through-extra-settling samples; zero samples outside the strict port region.
Normal N=16 16/16 PASS All 57 stages, 5,434 controls. Each row had 2,043 release-through-extra-settling samples; zero samples outside the strict region.
N=2, only row0's slip observation forced over limit after first pinch Survivor FAIL Row0 was rejected as intended. Healthy row1 completed 41 stages, then naturally truncated in align_tip_2_1 at control12,001, before final insertion/release.

The normal cases include the task's original 2-second settling interval plus 5 additional physical seconds holding the last robot command. Fresh live endpoint measurements, rather than sealed success, stayed within [0.5845,0.064,0.753]..[0.5875,0.0896,0.759] m throughout opening, retreat, home and extra settling. Maximum real slip across the 18 normally completed tasks was 2.019 mm, below12 mm. Maximum queried robot/fixture penetration was about0.027 mm, below1 mm; final sampled bench/spool/port penetration was0. No contact rows were dropped and no cross-arena robot/fixture contacts were queried.

Original first-spool and third-carry milestones also passed in the 16-row run. Resetting only row0 afterward left the other15 cable states and robot qpos/qvel/targets bitwise unchanged, with no physics-clock advance.

What is fixed

The added pitch/pivot arcs, grip adjustment, downward correction margin and release sequence resolve the prior insertion/release failure for these normal fixed-seed batches. The new per-row precondition outcome also removes the earlier direct batch-wide ValueError: after rejecting row0, row1 now proceeds through many later stages.

Remaining P2: inactive rows still determine subsequent trajectory durations

New inline finding. The cable lowerer calculates distance/angle maxima over the complete service batch. The inactive row no longer executes, but its held posture still affects those reductions. The compiler filters eligibility only after lowerer computations, so subsequent paths become much longer for the healthy row.

Actual same-stage control counts:

Stage Normal N=2 Row0 rejected, row1 healthy
align_over_port_0 46 649
align_over_port_1 42 649
align_over_port_2 21 649
align_over_port_3 20 649
lower_channel_0 20 343
align_tip_0_0 20 649

The rejected row's left TCP stays near its initial (0.45,0.26,0.93) m posture. maximum_y=0.05 then makes its computed, unexecuted target roughly0.21 m away. At0.02 m/s and60 Hz this contributes roughly630 Cartesian samples on every such call, plus normal hold/check controls. Because the row never executes the path, the same cost recurs at the next correction.

The native run retained the original max_episode_steps:12000. It ended naturally with:

last stage: align_tip_2_1 (42/57)
controls: 12001
terminal_reasons: [segment_validation_failed, truncated]
truncated: [false, true]
successful fragments: [2, 41]
healthy row1 final max slip: 0.421529 mm
healthy row1 sampled bench/spool/port penetration: 0

No physical-state corruption, solver modification or artificial stop was used. The only injected value was the first row's runtime cumulative-slip observation. An independent CPU probe with non-contiguous physical IDs (3,1) reproduces the full-service timing inflation using finite poses and compares it with a survivor-only service.

Reproduction

Use the copy-paste script already posted in the previous diagnostic feedback, saved as /tmp/pr834_parallel_repro.py. It supports the current57-stage program without changing production source. From a checkout of d2a182945b994d59ec302870286af812de6e2c39 with the qualified YAM asset and rc2 SDK:

export PYTHONPATH="$PWD:$PWD/extensions:$PWD/extensions/embodichain_tasks:$PWD/extensions/embodichain_models:$PWD/extensions/embodichain_benchmarks:${PYTHONPATH:-}"
python /tmp/pr834_parallel_repro.py --num-envs 2 --seed 73 --output /tmp/pr834-d2-normal2.json
python /tmp/pr834_parallel_repro.py --num-envs 16 --seed 73 --output /tmp/pr834-d2-normal16.json
python /tmp/pr834_parallel_repro.py --num-envs 2 --seed 73 --fault-row 0 --output /tmp/pr834-d2-survivor.json

The synthetic observation is applied only after the acknowledged right_pinch_0.162 stage:

runtime = env._active_task_program_bridge.post_policy_port._cable_runtime
runtime.max_slip[0] = 0.013

Keep the original episode limit and allow the healthy row to run to its normal completion/truncation. Read per-row terminal reasons, stage spans and success in the saved metadata; a successful process exit from an external diagnostic harness is not itself task acceptance. Physical contact variation can change exact counts; the failure criterion is unintended healthy-row truncation caused by inactive-row timing.

Correction and re-acceptance

  1. Propagate eligibility to the registered lowerer/runtime-service computations before aggregate timing and geometry decisions. Exclude rejected rows from the shared distance/angle maxima; preserve their held values in full-batch output.
  2. Add an independent finite-state CPU regression: changing only an inactive row's pose must not change active-row sample counts/durations. Cover reordered physical IDs and all-inactive behavior.
  3. Rerun the native early-slip case: row0 should fail, while row1 completes all57 stages within the original12,000-control budget and passes opening, retreat, 2s settling and5 additional physical seconds. Then repeat the normal2/16 cases.
  4. Keep the unified DexUni path and original strict region,12 mm slip and1 mm penetration criteria.

Review scope and limits

Three independent static/CPU passes covered robotics, architecture/runtime and verification. The material/VisualObject/package selection passed105 cases, selected action checks14, vertical slices20 and the existing visualization-coverage gate1. The independent geometry probe checked120 pivot arcs, with maximum pivot component error5.96e-8 m. These CPU results do not override the failed native survivor case.

The verified physical assets/material owners and evaluation/execution-policy files remain byte-identical to the earlier367fc038 failure baseline. The initial robot posture was changed before this reviewed delta; this run uses the latest posture and does not claim equality with the earlier initial state.

Hardware/runtime: RTX5090, Python3.12.16, Torch2.11.0+cu130, DexSim0.6.0rc2 build dbaca660471d5aaa7e750093e7fe1c207465c4af, seed73, NoRender. Each case ran in a fresh process, sequentially. This is targeted fixed-seed acceptance, not a multi-seed success-rate claim. Production source, native physical state and solver settings were not changed by the diagnostics. The remaining confirmed P2 should be corrected before treating the parallel survivor behavior as accepted.

Carry the net cable stack changes from ac38b4d relative to the already integrated main snapshot. Retain Task Program routing, contact grasping, USD cable lifecycle and override support, demos and regression coverage.
Check authored joint speeds from the observed pose through the last sample. Carry measured cable check failures through grounding and planning by physical environment, and verify that healthy rows can finish the following stage.

Reconcile the drawer declaration fingerprints and enable Viser updates in the two-post cable preview, with and without video.
Preserve correction travel, lower the held distal span before release, and withdraw vertically while retaining the existing physical acceptance gates.

Validate pinned rc2 with four dual-environment and three 16-environment batches, the original full-route and third-carry CUDA regressions, and 104 focused CPU tests.
Use the configured arm component before resolving the default asset so offline configurations round-trip. Keep the release task prerequisite aligned with the DexSim version pinned on main.
@KeqiYe
KeqiYe force-pushed the codex/cable-object branch from d2a1829 to 5517b73 Compare November 19, 2026 07:27

@han-xudong han-xudong left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Request changes for two P2 environment-isolation defects on 5517b73832cf564bba39d3e911f7159ac3d307d7:

  1. Batch-wide IK failure (simulation/cable.py:323): one environment with no reachable wrist branch aborts healthy environments before any command. The new inline comment in this review includes the production executor/bridge CPU reproduction.
  2. Inactive-row duration coupling (simulation/cable.py:603, also the rotation reduction): rejected environments still affect the shared trajectory duration. I added the independent 12-to-589-control reproduction to the existing inline discussion.

Please preserve healthy-row execution by reporting grounding failures per environment and applying eligibility before shared duration calculations.

Validation: 865 focused CPU cases passed, and the remaining case passed after correcting the local extension import path. Both findings were independently reproduced with CPU probes. These checks used DexSim 0.6.0rc1; native simulation under the PR's declared 0.6.0 was not rerun.

@yuecideng

Copy link
Copy Markdown
Contributor

Suggestion: halve the physical cycle time with phase-specific pacing

@KeqiYe, the timing experiment suggests that a roughly 44–45 second physical cycle is achievable, but the accelerated profile needs further contact-stability qualification before becoming a default. I recommend a separate task timing profile with faster transit motions and explicitly preserved critical contact-settling phases.

Evidence scope: the controlled experiment used d2a182945b994d59ec302870286af812de6e2c39, matching the immediately preceding baseline acceptance. The current 5517b73832cf564bba39d3e911f7159ac3d307d7 head has not been run with this experimental profile. No timing change was pushed to this PR.

Measured results

Case Baseline physical task time Accelerated time Completed tasks
2 environments 88.37 s / 5,302 controls 43.38 s / 2,603 controls 2/2
16 environments, first batch 90.57 s / 5,434 controls 44.35 s / 2,661 controls 15/16
16 environments, fresh repeat 90.57 s / 5,434 controls 44.12 s / 2,647 controls 16/16

This reduces physical task time by approximately 51%. The original final two-second settling interval remains inside these task times; the additional five-second persistence check is excluded from them. Successful rows retained their endpoints in the unchanged strict channel throughout release, retreat, home and the extra five physical seconds.

The experimental profile used:

  • Effective translational speeds for all 42 moving targets multiplied by 2.15, including the 13 targets that previously relied on the default speed.
  • Task-local angular rate increased from 0.5 to 1.075 rad/s.
  • PickUp sampling, gripper interpolation and process-hold frame counts shortened proportionally, retaining minimum control-frame constraints.
  • Planning acceleration budget increased from 0.5 to 2.31125, following the squared time-scaling relation. Original Robot joint velocity caps remained enforced.
  • Unchanged 60 Hz control rate, 1/240 s physics timestep, solver substeps/iterations, unified DexUni coupling, physical assets/materials, initial posture, geometric path/pivot constraints, strict port bounds, 12 mm slip limit and 1 mm penetration limit.

The action rate is slightly above 2 because the final acceptance hold and minimum command/check frames are retained. The measured total duration, rather than the parameter multiplier, defines whether the physical cycle was halved. This was an actual faster native simulation trajectory; the preview was rendered from its recorded physical states at simulation-time playback speed.

Observed limitation

In the first accelerated 16-row batch, row 10 was inside the channel after lowering and tip alignment (endpoint Z approximately 753.6 mm). After the second wrist lay-down, it rose to 759.874 mm, then 759.907 mm at verify_release_alignment, approximately 0.907 mm above the unchanged 759 mm upper bound. Release was correctly blocked for that row, while the other 15 rows completed.

Its maximum measured grasp slip was 1.175 mm, below 12 mm. The failure manifested as an endpoint rise following wrist lay-down. Since the experiment increased motion rate and shortened waiting together, it does not separately identify their contributions or prove that restoring one wait will fix every case. The subsequent 16/16 batch does not erase the earlier failure or establish equal reliability to the baseline.

Recommended next candidate

  1. Keep faster free-space approach, transit and home motions. Give linear and angular pacing explicit per-phase configuration so the timing can be evaluated without changing a common hard-coded angular constant.

  2. Restore the original settling around sensitive contact transitions first, while retaining the faster transit profile:

    • settle_soft_grasp: 30 controls instead of the experimental 14.
    • Both post-lay_down_free_end holds: 15 controls instead of the experimental 7. Use a dedicated contact-settle target rather than slowing every settle_motion occurrence.
    • settle_released_end: 30 controls instead of the experimental 14.

    This adds only 48 control frames, approximately 0.8 s, arithmetically. Relative to the measured 43.38 s two-row cycle, the nominal result would remain near 44.18 s, before feedback-dependent changes in subsequent trajectories. This candidate has not been executed and is not yet qualified.

  3. Use bounded, measured settling where appropriate. Continue holding the loaded command while observing actual endpoint/velocity stability and slip. Keep TCP tracking and actual flexible-tip convergence as separate conditions. Preserve the original final two-second acceptance settle and the extra five-second live-physics check.

  4. Retain the unified DexUni solver and existing physical criteria. The timing optimization should not require looser channel bounds, slip/penetration limits, a larger episode budget, physical-state writes or runtime attachments.

Re-qualification

  • Test the new profile on the current PR head in fresh 2-row and 16-row processes; record achieved physical duration and per-row outcomes.
  • Require complete insertion, opening, retreat, home, final settling and live endpoint persistence under the original strict bounds. Preserve intermediate spool milestones and robot/fixture contact checks.
  • Repeat batches and report actual successes/attempts; these fixed-seed observations are not a statistical success-rate claim.
  • Keep the original timing as a reference profile until the faster candidate has passed the same physical gates consistently. Include the early-rejected/healthy-row continuation test so timing changes do not hide the separate eligibility/timing coupling discussed in the earlier acceptance feedback.

The material/pivot/grasp CPU suite passed 70 cases for the experimental profile, and deployment preflight passed. Those checks support the configuration and path mathematics; the observed 15/16 native batch remains a failed full-cohort qualification.

@yuecideng yuecideng changed the title Add contact PickUp and Task Program cable routing Add eight-call cable routing with contact-aware fast pacing Nov 19, 2026
@yuecideng yuecideng added the performance Performance, memory, throughput and synchronization-cost improvements. label Nov 19, 2026
@yuecideng

Copy link
Copy Markdown
Contributor

Post-main validation follow-up on a5cd502ee: the canonical N=2 run passed 2/2 at 44.18 s and all three CUDA scenarios passed, but the final N=16 batch was 15/16, so it is not a full-cohort pass.

Row 11 passed the two final tip corrections at Z=753.60 mm. After wrist lay-down 0 it was at 754.58 mm; after lay-down 1 and its 0.25 s hold it reached 759.263 mm, above the unchanged 759 mm bound. release_alignment correctly rejected that row. Other 15 rows completed. With its loaded command held, row 11 relaxed below the bound roughly 0.65 s later, but never completed release/retreat; its eventual in-channel position therefore does not count as task success.

I am extending only the second lay-down's contact settling to one second, retaining its slow angular rate and all physical gates, with a small free-motion pacing adjustment to retain an approximately 44–46 s cycle. This candidate will be re-qualified in fresh 2/16-row processes before merging. The previous 34/34 observations in the description remain explicitly scoped to d9eff2f47; they do not erase this later failed batch.

@yuecideng
yuecideng dismissed han-xudong’s stale review November 19, 2026 15:44

The two requested row-isolation corrections on 5517b73 are implemented in a5cd502: eligible masks are carried into registered grounding, inactive rows are excluded from timing, and mixed IK/lift/ray failures return row-specific finite outcomes. Both discussions have detailed replies and are resolved. Independent engine/verification passes and the new CPU regressions confirm the corrections. The old review is superseded by the updated implementation; merging is being held for final physical qualification and current-head CI.

@yuecideng
yuecideng merged commit 763c5c7 into main Nov 19, 2026
6 of 8 checks passed
@yuecideng
yuecideng deleted the codex/cable-object branch November 19, 2026 15:53
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

assets Robot and scene assets, meshes, materials, asset catalogs and downloads. atomic-skill Atomic Skill goals, planning, execution, verification and recovery. breaking User-visible incompatible changes requiring migration. bug Incorrect existing behavior or a fix restoring the intended behavior. compute Shared numerical algorithms and kernels independent of simulation objects. docs Documentation, API references and example explanations. feature New user-facing capabilities. gym Environment lifecycle, managers, configuration composition and registration. kinematics FK, IK, Jacobians, reachability and kinematic solvers; not physics solvers. motion gen Robot motion planning, MotionGenerator, trajectory generation, retiming and expansion. performance Performance, memory, throughput and synchronization-cost improvements. physics Physics backends, dynamics, contacts, fluids, deformables and physics solvers. robot Robot runtime, presets, control groups, calibration and robot support. task Concrete tasks, deployments, objectives and recipes, including configured Task Program tasks. task-program Task Program language, semantics, compiler, runtime and integration contracts.

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants