Nam2452110/hm3d_2d

HM3DNav

2D indoor navigation Gymnasium environment built from real HM3D habitat floors, with per-task Signal Accessibility (SA) labels computed exactly via absorbing Markov chains (the Phase A "Signal Starvation" study).

What is inside

• hm3dnav.env.HM3DNavEnv — Gymnasium env
  • Action Discrete(4): UP / DOWN / LEFT / RIGHT
  • Observation: 22-ray LiDAR distances to walls
  • Sparse reward: +1 on reaching the goal, episode budget 200
  • reset() also returns p0, sa_uniform, sa_persist, starved, level
• hm3dnav/data/grids/ — 364 floor grids (*.npz, 0=free, 1=wall)
• hm3dnav/data/tasks/ — 364 task files (17 tasks per floor, 6178 total)
• hm3dnav/data/difficulty.json — per-task labels keyed by map id

Labels (Signal Starvation)

A task is starved when sa_uniform < 1e-4, i.e. a uniform random-walk policy is expected to reach the goal fewer than once in 1000 training episodes (1000 x sa_uniform < 1).

level	sa_uniform range	tasks	fraction
easy	>= 0.1	16	0.3%
medium	[1e-3, 0.1)	1272	20.6%
hard	[1e-4, 1e-3)	521	8.4%
starved	< 1e-4	4369	70.7%

These labels are identical to the Phase A experiment: 70.7% of HM3D tasks are starved (vs 0.8% on standard benchmarks).

Install

pip install -e .

Quickstart

from hm3dnav.env import HM3DNavEnv
from hm3dnav.tasks import sample_task, list_maps

env = HM3DNavEnv(map_id=list_maps()[0], task_idx=0)
obs, info = env.reset()
print(info["goal"], info["sa_uniform"], info["starved"], info["level"])
# obs: 22 LiDAR distances

Reproducing the labels

scripts/build_difficulty.py regenerates data/difficulty.json from p0_discrete/hm3d_full/phase_a/results/phaseA_probes.pkl (the Phase A probe measurement). Starved fraction comes out to 70.719%.