omx-act-sponge-bowl

OMX ACT: Place Sponge in Bowl

Custom ACT base policy for the six-DoF OMX follower task place the sponge in the bowl. It was exported at training step 80,000 and predicts 30 absolute joint targets per chunk.

Runtime contract

• Joint order: shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper
• Cameras: wrist, top
• Image shape: 3 x 480 x 640
• Control rate: 30 Hz
• Action space: six absolute joint targets
• Training dataset: RevanthGundala/014-place-sponge-in-bowl-diverse-50

Required code

The custom_act architecture and deployed evaluator are pinned in RevanthGundala/omx-act-recap at commit 90d460db2de189b7d7cf0a2d71c7b33b1d5eb705.

Clone ResFiT with its pinned submodule and download this complete model directory:

git clone --recurse-submodules https://github.com/amazon-far/residual-offpolicy-rl.git
cd residual-offpolicy-rl
hf download RevanthGundala/omx-act-sponge-bowl --local-dir pretrained_model

Run the unchanged deployed base-policy evaluator:

uv run python -m resfit.omx.evaluate_act \
  --policy-path "$PWD/pretrained_model" \
  --policy-type custom_act \
  --policy-device mps \
  --fps 30 \
  --actions-per-chunk 30 \
  --chunk-size-threshold 0.4 \
  --aggregate-fn conservative \
  --action-smoothing 0.7 \
  --max-joint-delta 40

Integrity

model.safetensors SHA-256: 81729758a3aa76f1696bc039f33b5f6ea0f827308bf0aa807a14fcdcf7ab9890

Use the saved preprocessor and postprocessor files with the weights. Loading only model.safetensors omits the training normalization and is not equivalent.

Hardware execution can move the robot. Verify the port, camera mapping, start pose, and emergency-stop procedure before evaluation.