vshwanilgv/wenavigatecontroller-long-episodes

WeNavigate — Controller Episodes

Expert navigation episodes collected from the HM3D minival dataset using GreedyGeodesicFollower inside Facebook Habitat-sim 0.3.1. Intended for behaviour-cloning pre-training and PPO fine-tuning of a discrete-action indoor navigation controller.

Dataset stats

Split	Episodes	Scenes	Total steps
train	160	10	9,955
eval	40	10	2,510

• Success rate: 100% (GreedyGeodesicFollower oracle)
• Geodesic distance: 1.5 – 15.0 m per episode
• Image size: 256 × 256 RGB
• Action space: 5 discrete actions (0=FORWARD, 1=LEFT, 2=RIGHT, 3=BACKWARD, 4=STOP)

NPZ schema (per file)

Array	Shape	dtype	Description
observations	(T, 256, 256, 3)	uint8	RGB frames
actions	(T,)	int32	Discrete action at each step
positions	(T, 3)	float32	Agent XYZ position
rotations	(T, 4)	float32	Agent quaternion (x,y,z,w)
episode_starts	(E,)	int32	Start index of each episode in T
episode_lengths	(E,)	int32	Step count per episode
scene_id	scalar	str	e.g. "00800-TEEsavR23oF"
split	scalar	str	"train" or "eval"

Loading example

import numpy as np

data = np.load("train/00800-TEEsavR23oF_train.npz", allow_pickle=True)

starts  = data["episode_starts"]
lengths = data["episode_lengths"]

for i, (s, l) in enumerate(zip(starts, lengths)):
    obs  = data["observations"][s : s + l]   # (T, 256, 256, 3)
    acts = data["actions"][s : s + l]         # (T,)
    print(f"Episode {i}: {l} steps")

Scenes (HM3D minival)

• 00800-TEEsavR23oF
• 00801-HaxA7YrQdEC
• 00802-wcojb4TFT35
• 00803-k1cupFYWXJ6
• 00804-BHXhpBwSMLh
• 00805-SUHsP6z2gcJ
• 00806-tQ5s4ShP627
• 00807-rsggHU7g7dh
• 00808-y9hTuugGdiq
• 00809-Qpor2mEya8F

Collection config

{
  "total_episodes": 200,
  "train_ratio": 0.8,
  "image_size": 256,
  "min_geodesic_dist": 6.0,
  "max_geodesic_dist": 15.0,
  "max_steps": 200,
  "success_dist": 0.2,
  "forward_step": 0.25,
  "turn_angle": 30.0,
  "seed": 42
}