An Efficient Learning-Based Task Planning Approach Using a Bio-Inspired Action Context-Free Grammar for Bimanual Manipulation
Carmona David, Jun Yang, Haoyong Yu
AI summary
Problem
Classical Task and Motion Planning suffers from combinatorial explosion, causing intractable planning times and latency that hinder real-world bimanual robot deployment and human-robot interaction.
Approach
The authors propose BAG-Learn Planning, which uses a Long-Short-Term Memory network to learn symbolic task plans from annotated human demonstration videos, encoding them via a Bio-Inspired Action Context-Free Grammar to replace traditional symbolic search with efficient sequence prediction.
Key results
- Planning times reduced from tens of seconds to under 100 milliseconds
- Sub-7ms latency maintained regardless of object count or locations
- Successful physical deployment for unseen pouring, opening, and passing tasks
- Compact 131k-parameter model enables embedded hardware deployment
Why it matters
Enables real-time, scalable task planning for bimanual robots, making human-robot collaboration more practical and responsive in household environments.
Abstract
Syntax Tree (AST). The interpreter iterates over the AST and derives a sequence of motion planning commands, which are executed using the RRT motion plan- ner. The TPEF enables the integration of the BAG-Learn ICRA2026 Late Breaking Results Poster presented at 2026 IEEE International Conference on Robotics and Automation (ICRA 2026) June 1-5, 2026. Vienna, Austria