paper_rob__ibrl

What problem does it solve?

IBRL provides a sample-efficient approach to robotic manipulation by dynamically selecting between a reinforcement learning (RL) policy and a behavior cloning (BC) policy using a dual-Q critic. This enables faster learning from demonstrations and online interactions, reducing data requirements and improving performance on complex tasks.

Core Features & Use Cases

• Q-based action selection between RL and BC actions for exploration and bootstrapping.
• Supports multiple modalities: pixel-based Robomimic tasks and state-based Meta-World tasks, with configurable encoders and critics.
• Training paradigms: IBRL, RLPD, and RFT, across both simulation and real-like settings, with pretraining and fine-tuning workflows.
• Real-world use case: a robot learns from demonstrations and gradually takes over with RL as it improves, reducing annotation costs.

Quick Start

Load the IBRL config and start a training run to observe improved sample efficiency.