PyTorch implementations of DA:PPO, DA:REPPO, and DA:WPO, with installation
instructions, multimodal behavior GIFs, and reproducible demo commands.
The assets below are the exact checkpoints used for the README's StackCube
and PushT demonstrations. Download both with:
uv run --locked --extra maniskill3 python tools/download_demo_checkpoints.pyThe downloader verifies SHA-256 checksums. See demo instructions
for matching configs, policy seeds, samplers, and training provenance.
Validation: 476 tests passed, 3 skipped, 33 subtests passed; Ruff passed;
CPU training smoke test and GPU training/replay checks passed. Both selected
StackCube modes and both selected PushT rotation directions succeeded.
The GIFs illustrate selected successful rollouts rather than aggregate success rates.