You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Access is granted manually. Please describe briefly how you plan to use this dataset.
Log in or Sign Up to review the conditions and access this dataset content.
OmniVLA Sim Go1 Navigation Dataset
Language-annotated robot navigation dataset in LeRobot v2.1 format, generated from a MuJoCo simulation of a Unitree Go1 carrying a ZED2 camera. Built to fine-tune OmniVLA (omni-modal vision-language-action navigation).
- 45 episodes / 7,724 frames at 10 Hz (~13 min), 640x360 RGB (h264)
- 16 scenes (indoor: hallway, office, lobby, warehouse, junction, doorway; outdoor: park, forest, orchard, trail, courtyard, campsite, construction, footpath, obstacle field, ramp)
- 91 language instructions (primary + paraphrases per episode) in
meta/tasks.jsonl/ episode task lists observation.state = [x, y, yaw]exact sim odometry;action = [v_lin, v_ang]executed velocity commands- Each episode is one collision-validated waypoint mission driven by a pure-pursuit controller (kinematic sim: odometry is ground truth)
Usage
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("adipotnis/omnivla-sim-go1")
For OmniVLA training, use the OmniVLASim_Dataset loader from the generating
pipeline: it derives OmniVLA's 8-step local-frame waypoint actions
[x/0.25, y/0.25, cos dyaw, sin dyaw] from observation.state at load time
and emits the train_omnivla.py sample dict (modality ids 4-8; language
available for every frame; no MBRA relabeling needed since actions are exact
odom).
Provenance
Scenes, mission definitions, and the full collection/conversion pipeline (MuJoCo bridge -> rosbags -> LeRobot) are code-generated; no real-world or human data is included. The Go1 robot model derives from mujoco_menagerie.
- Downloads last month
- 31