Learning robust and generalizable bimanual skills: a spatiotemporal causal hierarchical diffusion framework with...

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Bimanual visuomotor imitation learning allows robots to learn dual-arm skills from demonstrations.
• Existing methods face challenges in temporal synchronization, collision avoidance, long-horizon reasoning, and visual robustness.
• Diffusion-based policies struggle with long-term dependencies, spatial precision, and domain shifts.

Purpose of the Study:

• To develop a robust framework for bimanual visuomotor imitation learning.
• To address limitations in temporal synchronization, spatial precision, and robustness to visual distractions.
• To improve robot manipulation skills in complex environments.

Main Methods:

• Proposed the Spatiotemporal Causal Hierarchical Diffusion Imitation Learner (SCH-DIL).
• Integrated spatiotemporal hierarchical diffusion optimization for multi-scale action modeling.
• Incorporated causal visual representation learning and noise-robust diffusion modeling with attention anti-interference regularization.

Main Results:

• SCH-DIL consistently outperformed existing diffusion-based and imitation learning baselines on the RoboTwin 2.0 benchmark.
• Achieved higher success rates in both clean and domain-randomized inference settings.
• Demonstrated robustness to visual distractions and domain shifts.

Conclusions:

• The proposed SCH-DIL framework offers a practical and robust solution for bimanual visuomotor imitation learning.
• Achieved significant improvements with minimal computational overhead.
• Effective for visually challenging and domain-shifted robotic manipulation tasks.