Skip to main content
Back

George Mason Teaches an 8B LLM to Navigate Tight Doorways With 92.6% Success

A bi-level LoRA fine-tuning recipe teaches Llama-3.1-8B to plan rigid-body motion through sequential narrow doorways with 92.6% success.

Pro

Paper, Full Access
  • George Mason researchers fine-tune Llama-3.1-8B to plan rigid-body motion through sequential narrow doorways in 2D.
  • Bi-level pipeline: failure-driven LoRA SFT on human demos, then GRPO refinement.
  • Reward comes from a deterministic geometry verifier checking boundary, collision, swept-collision, and step-size constraints.
  • Achieves 92.6% in-distribution success and 77.3% out-of-distribution, versus 5.2%/2.1% for the base model.
  • Exit poses from one opening automatically bias toward the heading needed for the next passage.
  • Currently 2D only with symbolic scene input; 3D SE(3) and real-world evaluation planned as future work.

Getting a robot to shove a T-shaped object through one narrow doorway is hard. Getting it through two doorways in a row, where the pose you pick at the first door determines whether you can even fit through the second, is a classic long-horizon geometry problem that trips up both classical planners and learned models. A new paper from George Mason University proposes a surprisingly effective fix: fine-tune an LLM to output waypoints, then reward it for geometry that actually checks out.

The authors, Al Jaber Mahmud and Xuan Wang, describe the approach in Geometry-Aligned LLM Fine-Tuning. Their system takes a symbolic scene description in text (workspace bounds, obstacle rectangles, object vertices, start and goal poses) and returns a fixed-length SE(2) waypoint sequence that a downstream planner densifies into a continuous trajectory.

Overview diagram of the geometry-aligned LLM planning pipeline

Why sequential doorways break existing planners

The core difficulty is that the configuration used to traverse an early opening constrains the set of reachable configurations for subsequent ones. If a robot carrying a long object exits the first door with the wrong heading, the corridor between doorways may be too narrow to rotate before hitting the second one.

Existing tools each fall short in a different way:

  • Sampling-based planners like PRM and RRT-Connect struggle in thin regions of configuration space and require careful tuning of grid resolution, sampling budget, and step size.