openEuler/IB_Robot_ACT_banana_pick

ACT banana pick — 1 arm, 2 cameras (step 364000)

Original ACT policy checkpoint for the single-arm banana-pick task, plus its RK3588-compiled variant. Legacy flat layout (no v3 inference manifest) kept for traceability.

• Model type: ACT (Action Chunking Transformer), chunk size 100
• Training checkpoint: step_364000 (2025-05-15 run)
• Source dataset: 1arm_2cam_banana_pick_v1_20260514
• Validation: passed on the physical cell

Observations / actions

stream	shape
observation.state	6
observation.images.top	3×480×640
observation.images.wrist	3×480×640
action	6

The RKNN variant re-encodes the two cameras at 3×240×320 (hand_view / top_view) for RK3588 NPU execution.

Files

• pytorch_model/ — PyTorch ACT weights (model.safetensors, ~2.3 GB), pre/post-processor configs and normalizer safetensors, train_config.json
• rknn_model/ — act_ros2_rknn.rknn (~120 MB) compiled for RK3588, with its own config.json

Usage

Load the Torch checkpoint with LeRobot's ACT policy class and the pytorch_model/config.json feature layout. For the current IB-Robot v3-bundle deployment path (ascend / rknn / torch deployments routed through inference_manifest.json), use openEuler/IB_Robot_ACT_banana_pick_distill instead — it is the distilled successor of this checkpoint.