FARO Framework Uses LLMs and RL to Accelerate Robot Motion Planning

The high-dimensional, hybrid, and underactuated nature of humanoid loco-manipulation has long stymied fast planning of novel behaviors. In response, researchers propose FARO (Feasibility-Aware Robot Motion Optimization), a nested kino-dynamic framework that integrates rapid feasibility checking with dynamically consistent trajectory generation given a candidate contact sequence. By coupling this module with a feasibility-guided tree search and a Large Language Model (LLM)-based contact plan sampling strategy, the framework substantially improves the search process over conventional methods. The LLM injects semantic priors to prune the combinatorial space of contacts, while the kino-dynamic module ensures all generated trajectories respect robot dynamics and actuator limits.

Beyond planning, the team shows that the resulting trajectories are sufficiently high quality to be tracked by a reinforcement learning (RL)-based controller, bridging the gap between simulation and real-world execution. The paper demonstrates the approach on challenging loco-manipulation scenarios, where FARO generates feasible, smooth motions that can be deployed without additional fine-tuning. This integration of LLM-driven sampling, optimization-based feasibility checking, and RL-based control marks a significant step toward autonomous humanoid robots that can adapt to unseen tasks on the fly.