Download README.md from thelamapi/Next-Passage: direct link, hf CLI and curl.
- Browser
- Download file 3.35 kB
-
https://huggingface.co/thelamapi/Next-Passage/resolve/main/README.md
- Command line
-
hf download hf://thelamapi/Next-Passage/README.md
-
curl -L -o README.md https://huggingface.co/thelamapi/Next-Passage/resolve/main/README.md
license: apache-2.0
tags:
- planning
- robotics
- simulation
- next-lab
Next Passage
Next Passage routes a robot across a grid floor where doors open and close on a schedule and obstacles patrol corridors and randomly stall. The planner searches over space and time (with a wait action), so it can wait for a door or let an obstacle pass, and it replans whenever an obstacle deviates from its prediction. Simulation only.
Space: https://huggingface.co/spaces/thelamapi/Next-Passage — pick a scenario seed, obstacle density and stall probability; all three planners run in the browser.
Results (600 test episodes per planner: 100 seeds × 3 densities × 2 stall rates)
| Planner | Arrived | Episodes with a collision | Door closed on robot / episode | Steps ÷ static shortest length | Replans / episode | Search nodes / episode | Python ms / episode |
|---|---|---|---|---|---|---|---|
| Next Passage (time-aware) | 100.0% | 0.0% | 0.000 | 1.13 | 19.5 | 1438 | 10.9 |
| Static + reactive replanning | 100.0% | 50.0% | 0.173 | 1.26 | 1.7 | 434 | 1.3 |
| Static shortest path | 100.0% | 61.3% | 0.192 | 1.10 | 0.0 | 194 | 0.9 |
On the 600 episodes where every planner arrived, mean arrival steps were
30.8 (time-aware), 34.3 (reactive) and 30.1 (static).
The time-aware planner had 0 unsafe or unfinished episodes out of 600; they are listed in evaluation.json.
Share of episodes with any unsafe event (collision or door closing on the robot):
| Condition | Time-aware | Reactive | Static |
|---|---|---|---|
| low/stall=0.1 | 0% | 36% | 53% |
| low/stall=0.25 | 0% | 41% | 48% |
| medium/stall=0.1 | 0% | 64% | 74% |
| medium/stall=0.25 | 0% | 61% | 72% |
| high/stall=0.1 | 0% | 76% | 88% |
| high/stall=0.25 | 0% | 77% | 85% |
Replay video: one unselected test seed with all planners, then a failure case (static planner collision).
Method
- Grid 21×15, three walls with one open gap and one scheduled door each, pillars, 3/6/10 patrolling obstacles.
- Obstacles advance one cell per step along a back-and-forth loop and stall with probability p. They ignore the robot.
- Time-aware planner: A* over (cell, t) with moves {wait, N, S, E, W}, heuristic = static BFS distance (admissible), horizon 160 steps; cells predicted to hold an obstacle at t or t−1 are avoided (covers swaps and single stalls); doors are entered only when open on arrival. It replans when an observed obstacle phase differs from its prediction.
- Collisions: robot and obstacle in the same cell after a step, or swapping cells.
- The JavaScript port in the Space reproduces Python exactly (Node parity: 6 cases, 0 mismatches).
Usage
python inference.py --seed 20001 --density high --planner timed
python inference.py --selftest
Limitations
- Discrete grid, perfect localisation and perfect knowledge of the door schedule; obstacles are observed exactly.
- Obstacles never react to the robot; real crowds and vehicles do.
- Search effort grows with horizon and obstacle count; a real system would need a time budget per replan.
- Recorded on NVIDIA L4 (Python timing); no physical robot was used.

