m3/go2z1-grasp-vision-v2

Go2+Z1 Warehouse Grasp Vision Dataset (V2)

LeRobot v2.1-formatted dataset of 5000 scripted-IK Z1 arm grasp episodes with dual 224×224 RGB camera streams (base_cam mounted on the Go2 trunk top + wrist_cam mounted on the Z1 link06 end-effector).

Generated entirely in Isaac Sim 6.0.0.0 / Isaac Lab develop with random per-episode cube positioning. Designed as the training dataset for vision-conditioned VLA fine-tuning of NVIDIA GR00T N1.7-3B.

Schema

Key	Dtype	Shape	Description
observation.state	float32	[6]	Z1 joint positions (joint1 … joint6)
observation.images.base_cam	video (h264 yuv420p)	224 × 224 × 3 @ 20 fps	Trunk-mounted RGB (15° forward pitch)
observation.images.wrist_cam	video (h264 yuv420p)	224 × 224 × 3 @ 20 fps	Z1 EE-mounted RGB (70° FOV)
action	float32	[8]	eef_delta(6) + gripper.close + gripper.open
frame_index	int64	[1]	Step within episode (0-79)
timestamp	float32	[1]	Seconds since episode start
episode_index	int64	[1]	0-4999
task_index	int64	[1]	always 0
index	int64	[1]	global frame index
next.done	bool	[1]	end-of-episode flag

Language annotation:

{"task_index": 0, "task": "pick up the cube"}

GR00T modality config (NEW_EMBODIMENT) for this dataset:

"video":    ModalityConfig(delta_indices=[0],
                           modality_keys=["base_cam", "wrist_cam"]),
"state":    ModalityConfig(delta_indices=[0],
                           modality_keys=["z1_joint_pos"]),
"action":   ModalityConfig(delta_indices=list(range(0, 16)),  # 16-step horizon
                           modality_keys=["z1_eef_delta", "z1_gripper"]),
"language": ModalityConfig(delta_indices=[0],
                           modality_keys=["annotation.human.task_description"]),

Statistics

• Episodes: 5000
• Total frames: 400 000 (80 / episode)
• Total mp4 files: ~10 000 (2 cameras × 5000 episodes)
• Total parquet files: 5000
• Format: LeRobot v2.1 (codebase_version: v2.1)
• Compressed size: ~140 MB (efficient h264 encoding of small 224×224 frames)
• fps: 20

Layout

go2z1-grasp-vision-v2/
├── meta/
│   ├── info.json          # codebase_version=v2.1, features schema, splits
│   ├── episodes.jsonl     # one line per episode (length, language)
│   ├── tasks.jsonl        # {"task_index":0, "task":"pick up the cube"}
│   ├── modality.json      # GR00T modality index ranges
│   ├── stats.json         # action / state stats
│   └── relative_stats.json
├── data/
│   └── chunk-{000..004}/
│       └── episode_{000000..004999}.parquet
└── videos/
    └── chunk-{000..004}/
        ├── base_cam/episode_*.mp4
        └── wrist_cam/episode_*.mp4

Usage

from datasets import load_dataset

# Episode-level metadata
import json
with open("meta/info.json") as f:
    info = json.load(f)
print(info["features"])

# Per-episode parquet
import pyarrow.parquet as pq
ep = pq.read_table("data/chunk-000/episode_000000.parquet").to_pandas()
print(ep.columns)
# observation.state, action, frame_index, timestamp, episode_index, task_index, index, next.done

For training GR00T on this dataset, use the launcher we've open-sourced:

python -m gr00t.experiment.launch_finetune \
  --base-model-path nvidia/GR00T-N1.7-3B \
  --dataset-path /path/to/go2z1-grasp-vision-v2 \
  --embodiment-tag new_embodiment \
  --modality-config-path go2_z1_warehouse/stage3_gr00t_finetune/modality_config_z1.py \
  --tune-visual --tune-projector --tune-diffusion-model --no-tune-llm \
  --shard-size 64 --dataloader-num-workers 2 \
  --max-steps 30000 --global-batch-size 16

Full collection script: v2_pipeline/stage2_collect_vision.py.

Generation pipeline

Each episode runs a 3-phase scripted controller in Isaac Sim:

1. Reach (steps 0-29) — Z1 interpolates from Z1_HOME ([0,0,-0.005,-0.074,0,0]) to Z1_REACH ([0,1.5,-1.0,-0.54,0,0])
2. Grasp (steps 30-34) — gripper.close = 1
3. Lift + retract (steps 35-79) — interpolate back to Z1_HOME, gripper.open = 1 at t=65

Per-episode randomization: cube position (x ∈ [0.40, 0.60], y ∈ [-0.15, 0.15], z = 0.54).

Reference policies

Use this dataset for	Recommended base model
Vision-conditioned grasp	nvidia/GR00T-N1.7-3B
Trained companion (V2)	m3/go2z1-grasp-gr00t-vision-v2 (training in progress)
State-only baseline (V1, dummy video)	m3/go2z1-grasp-gr00t-n17-v1

Source code

• Repo: https://github.com/aws300/go2_z1_warehouse
• Collection script: v2_pipeline/stage2_collect_vision.py
• Modality config: stage3_gr00t_finetune/modality_config_z1.py

Citation

@misc{go2z1-grasp-vision-v2,
  title  = {Go2+Z1 Warehouse Grasp Vision Dataset (V2)},
  author = {m3},
  year   = {2026},
  url    = {https://huggingface.co/datasets/m3/go2z1-grasp-vision-v2}
}