openEuler/IB_Robot_ACT_dual_arm_banana_pick

ACT banana pick — dual arm (3 cameras)

Dual-arm ACT policy checkpoint for the banana-pick task with its Ascend OM compile and short verification clips. Legacy flat layout (no v3 inference manifest).

• Model type: ACT (Action Chunking Transformer)
• Source dataset: dual_arm/yzh/yzh_0606

Observations / actions

stream	shape
observation.state	12 (6 per arm)
observation.images.top	3×240×320
observation.images.left	3×240×320
observation.images.right	3×240×320
action	12

Files

• pytorch_model/ — PyTorch ACT weights (model.safetensors, ~2.3 GB), processor configs, train_config.json, and a Torch-side verification video
• om_model/ — Ascend OM compile (model.om, ~923 MB) with config.om.json (schema v1, ascend_om backend) and an OM-side verification video

Usage

Load the Torch checkpoint with LeRobot's ACT policy class. The OM artifact targets the Ascend (310P/310B) deployment of the same policy. For the current IB-Robot v3-bundle deployment path, see openEuler/IB_Robot_ACT_banana_pick_distill (single-arm distilled variant).