ACT Dual-Arm Vanilla (Baseline)
This model is a baseline Action Chunking with Transformers (ACT) policy trained for cooperative dual-arm manipulation in PyBullet. It was trained without domain randomization to establish a performance baseline on a single object (cracker_box).
Model Details
- Architecture: Action Chunking with Transformers (ACT) with a ResNet18 vision backbone.
- Task: Dual-arm cooperative lifting and placing of an object.
- Action Space: 12-D absolute joint angles in radians (6 degrees of freedom per arm, gripper state fixed).
- Vision: 3 camera streams (overhead, left wrist, right wrist) at 224x224 resolution.
- Training Data: 149 expert demonstrations collected at 20 FPS.
Performance
Achieved a 95% success rate and 100% grip rate on the in-distribution evaluation suite. The model exhibits highly deterministic chunk execution and low positional jitter compared to diffusion-based alternatives.
