cagataydev/scout-earth-rover-mini-20260616-053232

scout — Earth Rover Mini · Apartment Tour (2026-06-16)

LeRobot v3 dataset captured from scout, an Earth Rover Mini sidewalk robot, during an indoor apartment exploration session. Each episode corresponds to one natural-language instruction given to the agent, with synchronized front+rear camera video, full telemetry state, and the action stream issued by the high-level policy (an AWS Strands agent driving the robot via the Earth Rover SDK).

🏠 The robot was asked to autonomously map and navigate a residential apartment, learning the spatial structure across multiple turns of feedback.

At a glance

Field	Value
Robot	Earth Rover Mini
Codebase	LeRobot v3
Episodes	3
Frames	4,176
FPS	4
Duration	~17 minutes total
Cameras	front + rear (224×224 RGB, h264)
Audio	per-episode .wav sidecars (mono PCM16)
Action dim	3 (linear, angular, lamp)
State dim	16 (battery, signal, gps, imu, rpms…)
Size	~111 MB
Recorded	2026-06-16 05:32 UTC

Episodes

#	Length (frames)	Task
0	1,677	Initial autonomous apartment tour — discover spatial structure, store in memory.
1	1,748	Navigate behind the operator to find the bedroom (avoid the bathroom).
2	751	Correction — the bedroom is in the opposite corner; retry navigation.

The task strings show the conversational refinement loop: scout attempted the mission, the operator corrected it, scout retried. This makes the dataset useful for studying multi-turn instruction following and spatial grounding from egocentric video.

Schema

{
    "observation.images.front": video[H=224, W=224, C=3],   # front cam, h264 mp4
    "observation.images.rear":  video[H=224, W=224, C=3],   # rear cam,  h264 mp4
    "observation.state":        float32[16],                # telemetry vector
    "action":                   float32[3],                 # [linear, angular, lamp]
    "timestamp":                float32,
    "frame_index":              int64,
    "episode_index":            int64,
    "task_index":               int64,
}

Action vector — normalized commanded twist plus lamp state:

• action[0] linear velocity (forward/back), in [-1, 1]
• action[1] angular velocity (yaw rate), in [-1, 1]
• action[2] lamp on/off (0.0 or 1.0)

Loading

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("cagataydev/scout-earth-rover-mini-20260616-053232")
print(ds)                       # 3 episodes, 4176 frames
print(ds[0])                    # first frame: dict with all features
print(ds.meta.tasks)            # list of natural-language tasks

To load a specific episode:

ep0 = ds.hf_dataset.filter(lambda x: x["episode_index"] == 0)

Recording setup

• Hardware: Earth Rover Mini (frontiers.ai sidewalk robot platform)
• SDK: earth-rovers-sdk on localhost:8001, polled at 4 Hz for video/telemetry
• Agent: AWS Strands Agent (Bedrock Claude) issuing high-level commands; per-frame action vector is the last commanded twist + lamp.
• Recorder: tools/_recorder_engine.py from cagataycali/earth-rover-mini — captures front + rear simultaneously via a thread pool, resizes to 224×224, encodes to h264, writes LeRobot v3 chunked parquet/mp4.
• One episode = one user turn. The operator gives an instruction, the agent acts, the agent ends the episode when the task is done (or aborted).

Intended uses

• Train Vision-Language-Action (VLA) policies on natural-language indoor navigation instructions.
• Study multi-turn task correction — episodes 1 and 2 are the same navigation goal with operator feedback in between.
• Egocentric front+rear indoor mapping / SLAM benchmarks.
• Action-chunking baselines (paper ref: arXiv:2601.09444v2 — H=10 chunks @ 4 Hz).

Notes & caveats

• Indoor only — the Earth Rover Mini is a sidewalk robot, but this session is fully indoors (apartment). GPS in observation.state is therefore unreliable / stale.
• No reward / success label — task success is implicit in the operator's next instruction (corrections imply failure, "good" implies success).
• Audio stored as audio/episode_NNNNNN.wav sidecars; LeRobot v3 does not yet model audio in the schema, but it's perfectly aligned to the video FPS.
• Single-operator dataset — small scale (3 episodes) intended as a probe / proof-of-concept, not a production training set.
• Privacy: the address mentioned in episode 0's task string has been left in the raw task strings as recorded by the agent; consumers may want to redact meta/tasks.parquet before public redistribution.

Citation

@misc{scout_earth_rover_mini_2026,
  author    = {Cagatay Cali},
  title     = {scout — Earth Rover Mini Apartment Tour Dataset},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/cagataydev/scout-earth-rover-mini-20260616-053232}
}

License

Apache-2.0 (data + code).