Programming-by-Demonstration for Long-Horizon Robot Tasks
summary
PROLEX synthesizes robot control programs by turning demonstrations into regex-like sketches, completing holes with an LLM-guided search, and pruning impossible partial programs via over-approximating regex pruning. The implementation targets long-horizon tasks with nested loops and conditionals, runs on a DSL, and solves 80% of 120 benchmarks within 120 seconds, recovers the ground-truth program from one demonstration for 81% of solved tasks, outperforms CVC5 (25% with a sketch) and beats GPT-Synth (0% on these environments).
AI-written summary · about
chapters (7)
- 0:00Programming-by-demonstration intro
- 2:47Motivation for robot policies
- 6:38Prolex algorithm overview
- 9:05LLM-guided sketch completion
- 11:09Pruning with abstract interpretation
- 12:56Evaluation results
- 14:01Q&A
related talks
15mAbstraction Refinement-guided Program Synthesis for Robot Learning from Demonstrations
25mHigh-Level Mission Specification for Multiple Robots
18mRobotics: why now?
16mInteractive Synthesis of Temporal Specifications from Examples and Natural Language
17mBeyond Automation: AI and the Evolution of Robotic Intelligence
09mSituated Live Programming for Human-Robot Collaboration
29mSynchronous Programming and Refinement Types in Robotics: From Verification to Implementation
16mKeynote: Robot Foundation Models