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
• hm3dnav/trainer/ — RL trainer (CLI + campaign driver) used by the cross-algorithm / cross-level comparison, with turn-and-thrust action models

RL contract and action models

The trainer runs on a different contract than the public HM3DNavEnv — the p0-validated RealMapEnv:

• Observation (22,): 16-ray egocentric LiDAR (/6 m) + relative goal (goal - agent, normalised by max(grid.shape); body-frame under turn models) + 4-dim last-action one-hot.
• Reward: sparse +1 on goal, episode budget 200.
• --action-model selects the motion model (all share the sparse goal reward):

model	action space	headings	step
cross4	Discrete(4) up/down/l/r	—	1 cell N/S/E/W
turn8	Discrete(3) l/fwd/r	8 x 45 deg	1 cell along heading
turn16	Discrete(3) l/fwd/r	16 x 22.5 deg	1 cell along heading

• Egocentric sensing: for turn models ray 0 = forward heading; relative goal is expressed in the agent body frame.
• Turn-aware distances: demo_env._turn_dist runs BFS over (cell, heading) — every primitive (a 360/n turn or a forward step) costs one step, so d_bfs / phi / SPL stay coherent with the Discrete(3) action space.
• --recompute-sa: re-derives p0 / sa_uniform / starved on the action graph (exact absorbing Markov chain over (cell, heading), trainer.selection.sa_for_model) instead of reusing the cross4 difficulty label. Turn models are harder for a uniform random policy (extra turn primitives), so SA drops vs the cross4 label.
• The continuous twin RealMapContEnv (Box(-1, 1, (2,)): heading turn + thrust) is used by SAC / SAC-HER and ignores the parameter.

RL trainer

Run from the repo root:

python -m hm3dnav.trainer --list
python -m hm3dnav.trainer --algo ppo --levels easy,hard --per-level 3 \
    --seeds 1 --steps 20000 --out results_trainer
python -m hm3dnav.trainer --algo ppo --map-id <map_id> --task-idx 0 \
    --action-model turn16 --recompute-sa --steps 5000 [--gui] [--gui-dt 0.01]

Per-cell output lands in {out}/{algo}_{level}[_{action_model}]/m{map_id}_t{task_idx}_s{seed}/ with summary.json, eval_curve.json and mech.csv; summary.csv aggregates all cells including the action_model and turn-aware d_bfs columns.

Trainer dependencies (stable-baselines3, torch) are extra:

pip install -e ".[trainer]"

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%.