Abstraction Refinement-guided Program Synthesis for Robot Learning from Demonstrations
summary
RoboScribe automatically derives state predicates and abstract actions from raw unsegmented demonstrations, alternates comparative abstraction refinement with program synthesis, and grounds abstract actions with neural low-level policies trained via standard RL. It learns subgoal predicates (e.g., hold, object-at-goal), fills loop guards, and produces iterative DSL programs that generalize 5→9 blocks in stacking tasks, showing a sharp success-rate jump after refinement and outperforming prior baselines; artifacts and code are available.
AI-written summary · about
chapters (10)
- 0:00Robot learning from demos
- 1:25Synthesis vs deep RL
- 3:53DSL for robot tasks
- 5:39Handcrafted abstractions limits
- 7:06Initial abstraction in RobotScape
- 8:12Abstraction refinement process
- 10:22Refining with subgoal predicates
- 12:26Learning loop guards and ordering
- 13:09Experimental results
- 14:15Conclusion and future work
related talks
21mProgramming-by-Demonstration for Long-Horizon Robot Tasks
24mReward-Guided Synthesis of Intelligent Agents with Control Structures
01h 10mNeurosymbolic Program Synthesis: Bridging Perception and Reasoning in Real-World(…)
01h 35mDeep Robotic Learning
18mActive Learning for Neurosymbolic Program Synthesis
01h 00mBetter Learning through Programming Languages
38mProgram Synthesis: A Dream Realized?
18mLearning Abstraction Selection for Bayesian Program Analysis