Skip to content

Releases: sphere-rl/sphere

HumanoidBench CRL SPHERE reproduction records (2026-05-04)

04 May 17:00

Choose a tag to compare

Run-record bundle for the public official-code sanity check of Top-K MoE PPO + SPHERE on the five-task HumanoidBench CRL setting.

WandB was disabled for these runs. The archive contains TensorBoard event files, raw evaluations.npz arrays, per-seed logs, run contracts, and aggregate summaries for seeds 0-4 on h1_stand, h1_walk, h1_pole, h1_slide, and h1_run.

The .sha256 asset verifies the archive.