Shapley-VLA — RoboTwin legacy checkpoints (Domain A, superseded)
The earlier RoboTwin-based Domain A of Shapley-VLA, kept because it holds the only converged multi-agent-RL-fine-tuned VLA weights the project has produced.
Domain A has since moved to RoboFactory; this domain is legacy and its code is preserved for reference rather than for reproduction.
Contents
| Path | What |
|---|---|
mix17_sft_step126637/ |
17-task mixed SFT — the shared parent of all five MAPPO runs |
mappo_grab_roller_step135/ |
MAPPO final |
mappo_pick_diverse_bottles_step270/ |
MAPPO final |
mappo_place_object_basket_step180/ |
MAPPO final |
mappo_handover_mic_step220/ |
MAPPO final |
mappo_place_bread_skillet_step235/ |
MAPPO final |
Weights only (full_weights.pt); FSDP dcp_checkpoint/ resume state is not uploaded.
⚠️ Normalization statistics were missing and have been restored
The MAPPO actor directories originally contained no norm_stats sidecar — only
dcp_checkpoint/ and model_state_dict/. The matching norm_stats.json has been copied
into physical-intelligence/robotwin/ inside each upload. Without it these weights emit
unnormalized actions: silently wrong, never an error.
Known result
Of the five tasks, RL lift was real on two (grab_roller 1.0, pick_diverse_bottles
0.31 → 0.63) and absent on three. Root cause was diagnosed as a pure sparse success reward:
RL can amplify a non-zero SFT base but cannot bootstrap tasks where SFT success is ~0.