Advances in Robot Imitation Learning

The field of robot imitation learning is moving towards more efficient and autonomous methods for adapting policies to new tasks and environments. Recent research has focused on developing techniques that can improve policy performance with minimal human intervention, such as using reinforcement learning to fine-tune policies learned from human demonstrations. Another area of focus is on improving the robustness and generalizability of imitation learning methods, including the use of knowledge-driven approaches and semantic keypoint graphs to abstract object representations. Notable papers in this area include: Steering Your Diffusion Policy with Latent Space Reinforcement Learning, which proposes a method for efficient autonomous adaptation of policies via reinforcement learning in latent space. SafeMimic, which presents a framework for safe and autonomous learning of mobile manipulation skills from a single human video demonstration. Human2LocoMan, which introduces a cross-embodiment imitation learning system for quadrupedal manipulation. CUPID, which proposes a data curation method based on influence functions to improve policy performance. DemoDiffusion, which enables one-shot human imitation using pre-trained diffusion policies. IMA-Catcher, which proposes an impact-aware framework for robotic catching of flying objects. Knowledge-Driven Imitation Learning, which leverages external structural semantic knowledge to abstract object representations and improve generalization capability.

Sources

Steering Your Diffusion Policy with Latent Space Reinforcement Learning

SafeMimic: Towards Safe and Autonomous Human-to-Robot Imitation for Mobile Manipulation

Human2LocoMan: Learning Versatile Quadrupedal Manipulation with Human Pretraining

CUPID: Curating Data your Robot Loves with Influence Functions

Robust Behavior Cloning Via Global Lipschitz Regularization

DemoDiffusion: One-Shot Human Imitation using pre-trained Diffusion Policy

IMA-Catcher: An IMpact-Aware Nonprehensile Catching Framework based on Combined Optimization and Learning

Knowledge-Driven Imitation Learning: Enabling Generalization Across Diverse Conditions

Built with on top of