Natural Language-based Robotic Planning and Manipulation
Object Detection, Planning with Large Language Models, Plan Execution and In-Hand Pose Estimation
This project aims to develop a robotic pick-and-place system that generates plans in natural language and executes object placement in specific poses, addressing complex constraints found in industrial assembly tasks. A key task example involves placing objects by specified shape and color into corresponding containers. The system is divided into detection, planning, and in-hand perception and control modules. These work in tandem to identify object properties, generate step-by-step plans from natural language inputs, and adjust the robot arm’s movement to ensure objects are correctly oriented to match container specifications. Each module has been integrated into a complete proof-of-concept system that demonstrates the feasibility of visual reasoning for pick and place, including in-hand perception to account for uncertainty in the attained grasp configuration.
The planning system uses Large Language Models while performing tree-search over actions from the paper SayCanPay: Heuristic Planning with Large Language Models using Learnable Domain Knowledge.