Shiki42/parallelvla-act-putcab-gripper-reweight-20260731

ParallelVLA ACT-small gripper reweighting experiment

Public deployment artifacts for a paired ACT-small experiment on the real RoboTwin 2.0 put_object_cabinet task.

Variants

• —uniform/: body coordinate weight 1, gripper coordinate weight 1.
• —gripper2x/: body coordinate weight 1, gripper coordinate weight 2.

Each directory is directly usable as RoboTwin ACT's ckpt_dir and contains:

• —policy_last.ckpt
• —dataset_stats.pkl
• —model_config.json
• —provenance and offline inference receipts

The optimizer/RNG training_state.pt files are intentionally excluded because this repository is the deployment payload, not the full-state training archive. The included source_checkpoint_manifest.json records the complete source checkpoint.

Model and data protocol

• —ACT-small, 21,884,431 parameters
• —ResNet-18 visual backbone
• —hidden dimension 256; FFN dimension 1024
• —4 encoder and 4 decoder layers
• —action horizon 50; action/state dimension 14
• —gripper indices [6, 13]
• —camera order head, left wrist, right wrist
• —deterministic training seed 87431
• —3,000 optimizer steps, batch size 8
• —40 training episodes and 10 held-out episodes
• —action alignment state[t] -> action[t+1]
• —chunks stop at phase or active-arm-mask boundaries
• —padded coordinates are excluded and losses are normalized by valid coordinate weight

Training dataset: robotwin_put_obj_cabinet_50_dynFcam_nFov-6280b6f02751, 50/50 native-success RoboTwin episodes and 10,464 frames. Wrist-camera geometry is centered_fovy90, with calibration manifest SHA-256 bd7a9d918a775f3d59b70ca04d1a9f4151c9602557487947e2ee49d8a461b0e5.

Source revisions:

• —parallelVLA training: 9585228ee63c855f8e6ccb55e794cf0eeba72954
• —parallelVLA ACT evaluation adapter: 095b7153b02808a5108a4306d7a4c851493f51fe
• —RoboTwin: c3ddfa8b97d5519efa828b075999bd0006778e5e

The evaluation/ directory contains the ACT algorithm configuration and the ACT/camera patch. Apply it after the repository's shared 0016-unified-official-evaluator.patch and 0020-make-policy-rng-modes-distinct.patch. The patch adds only ACT episode instrumentation, the training-matched wrist-camera installation/receipt, and parallelvla_eval_pcd.yml; it does not change the policy action loop.

Paired diagnostic result

All 3,000 sampled batches have identical sample/action/padding SHA-256 values across variants. On the 2,051 held-out observations, using ACT's deployment-time prior path:

This is a Pareto trade-off, not evidence that either policy has better task success. Native fixed-seed RoboTwin rollout evaluation is required for that conclusion.

Integrity

Use SHA256SUMS and download this repository at the immutable Hugging Face commit recorded by the experiment receipt. Do not deploy from a moving main revision.