Generalization in Robotic Behavior Cloning: Pitfalls and Promises
summary
Shows that adding and training on input noise — called replica noising — enforces total variation continuity so a conditional generative policy avoids exponential compounding in contractive systems and yields rollout error bounded roughly linearly in horizon. Algorithm: train conditional diffusion on noisy states and deploy by adding the same noise. Implements diffusion policy optimization in simulation and runs zero‑shot on real robot hardware: diffusion models remain robust while mean‑predictors fail.
AI-written summary · about
chapters (12)
- 0:00Introduction and context
- 2:28Defining behavior cloning
- 3:46Challenges and hypotheses
- 7:42Theoretical problem definition
- 10:46Understanding error accumulation
- 16:06Mitigating compounding error
- 18:30Total variation continuity
- 22:04Replica noising techniques
- 27:47Future directions and biases
- 30:47Practical applications
- 38:43Performance and exploration
- 43:12Conclusion and Q&A
related talks
01h 09mLearning Generative Models from Corrupted Data
42mReinforcement Learning using Generative Models for Continuous State and Action Space Systems
01h 21mGenerative AI with Diffusion Models
30mRepetita Non Iuvant: Why Generative AI Models Cannot Feed Themselves
01h 35mDeep Robotic Learning
54mGeneralization and Robustness in Offline Reinforcement Learning
37mMultiagent Reinforcement Learning: Rollout and Policy Iteration
41mBetter Learning from the Past: Counterfactual / Batch RL