Atharva Bhorpe · Project
Tiny Recursive Model for Path Planning
A compact recursive model that learns navigation policies on occupancy grids.
- Role
- Model development and evaluation
- Tools
- PyTorch · Transformers
Problem and objective
Learn the next action at each free cell of an occupancy grid, given a start and a goal. BFS/Dijkstra supplies the ground-truth policy during data generation.
My contribution and approach
- Implemented a recursive transformer for per-cell actions: north, south, east, west and stay.
- Used weight-shared recursion to increase effective depth without adding distinct layers.
- Evaluated path success, route optimality and latency against A* and parameter-matched CNN baselines.
The public README describes the model as approximately 1.2M parameters. The contribution summary is supported by the supplied résumé and repository documentation.
Reported results and limitations
- The repository reports 100% path success and optimal routes on 26×26 grids at 25% obstacle density.
- On 40×40 grids, it reports success rising from 52% to 75% with more test-time recursion.
- Larger-grid success remains below 100% in these reported tests.
These are source-reported grid benchmarks. The evaluations have not been independently reproduced for this portfolio. They do not establish performance on a physical robot.
Lessons
The reported larger-grid evaluations show a limit to generalization. More test-time recursion improves success in those tests, but does not eliminate failures.