cagataydev/kimodo-g1-motion-v1

Kimodo → Unitree G1 Motion Dataset (v1)

10,000 episodes of text-conditioned humanoid motion for the Unitree G1, generated by streaming nvidia/Kimodo-G1-RP-v1 diffusion into a MuJoCo physics rollout via strands-labs/robots.

• Robot: Unitree G1 (29 DoF, 65-dim proprioceptive state)
• Format: LeRobot v3.0 (parquet)
• Total frames: 1,990,000
• Total episodes: 10,000
• Episode length: 199 frames (4.0 s @ 50 Hz)
• Control frequency: 50 Hz
• Physics: MuJoCo (1 kHz sim, 50 Hz control)
• Motion source: Kimodo text-to-motion diffusion (kinematic prior)
• Tracker: ProtoMotions GTP (BeyondMimic) ONNX PD policy
• Task prompts: 25 base motion categories × ~400 stylistic variants ≈ 10 k unique prompts

Task categories (25)

idx	task
0	perform: a person raising both arms overhead
1	perform: a person bowing forward
2	perform: a person throwing boxing punches
3	perform: a person clapping hands with enthusiasm
4	perform: a person crouching down low
5	perform: a person dancing to music with rhythm
6	perform: a person jumping up with both hands raised
7	perform: a person performing a front kick
8	perform: a person sitting cross-legged in a meditation pose
9	perform: a person crouching down to pick up an object
10	perform: a person pulling something toward the body
11	perform: a person pushing a heavy object forward
12	perform: a person reaching forward with both hands
13	perform: a person running forward at moderate pace
14	perform: a person raising the right hand in a salute
15	perform: a person stepping sideways
16	perform: a person doing bodyweight squats
17	perform: a person performing bodyweight squats
18	perform: a person swaying gently side to side
19	perform: a person standing in a T-pose with arms out
20	perform: a person turning around in place
21	perform: a person twisting the torso left and right
22	perform: a person walking forward with confident strides
23	perform: a person walking backward carefully
24	perform: a person waving hello with the right hand

Episodes are approximately evenly distributed across the 25 tasks (~400 episodes/task).

Schema

Each row is one 50 Hz control step:

feature	dtype	shape	description
observation.state	float32	(65,)	G1 proprioceptive state (root pos/quat/lin+ang vel + 29 joint pos + 29 vel)
action	float32	(29,)	29-DoF PD joint-position targets (normalized)
timestamp	float32	(1,)	seconds from episode start
frame_index	int64	(1,)	step within episode (0..198)
episode_index	int64	(1,)	global episode id (0..9999)
task_index	int64	(1,)	task category id (0..24), maps to meta/tasks.parquet
index	int64	(1,)	global frame id

Sample row

observation.state[:5] = [-0.0114, -0.0054,  0.7768,  0.9999, -0.0112]
                        #  root_x   root_y  root_z   quat_w   quat_x ...

action[:5]            = [ 0.0991,  0.1640,  0.1117,  0.1538,  0.1554]
                        #  left_hip_pitch  right_hip_pitch  ...

Loading

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("cagataydev/kimodo-g1-motion-v1")
print(ds)                       # 1_990_000 frames, 10_000 episodes
sample = ds[0]                  # {"observation.state": ..., "action": ..., "task": "..."}

# By episode
ep = ds.hf_dataset.filter(lambda x: x["episode_index"] == 42)

# By task category
walk_forward = ds.hf_dataset.filter(lambda x: x["task_index"] == 22)

Streaming works out of the box:

from datasets import load_dataset
ds = load_dataset(
    "cagataydev/kimodo-g1-motion-v1",
    streaming=True, split="train",
)
for sample in ds.take(10):
    print(sample["episode_index"], sample["frame_index"])

Generation pipeline

text prompt
   │
   ▼
Kimodo diffusion (nvidia/Kimodo-G1-RP-v1)
   │   • 100 denoising steps, 30 Hz kinematic motion
   ▼
SLERP upsample 30 Hz → 50 Hz  (120 → 199 frames)
   │
   ▼
ProtoMotions GTP (BeyondMimic tracker, ONNX)
   │   • text-conditioned kinematic prior → PD joint targets
   ▼
MuJoCo physics rollout (1 kHz sim, 50 Hz control)
   │
   ▼
LeRobot v3 parquet shards (streamed to HF)

Per-episode timings on Thor (Jetson AGX Thor Dev Kit, 128 GB):

• Kimodo motion gen: ~8 s
• ProtoMotions FK conversion (cached after first): ~1 s (38 s first-run text-encoder init)
• Physics sim + record: ~1.2 s (3.2× realtime)
• Total: ~10 s/episode, or ~28 h wall for 10 k episodes.

Prewarmed qpos caches for all 25 base motions live in qpos_cache/ in the generator repo (≈ 5 MB total). This is what enables the streaming pipeline to hit ~10 s/episode.

Reproducibility

• Motion prior: nvidia/Kimodo-G1-RP-v1
• Tracker: ProtoMotions GTP unified_pipeline.onnx (uploaded to cagataydev/protomotions-gtp-g1)
• Robot model: Unitree G1 MJCF from strands-labs/robots
• Generator code: internal @ cagataycali/robots fork, branch feat/kimodo-mujoco-integration (PR ↔ strands-labs/robots pending)
• Prompt schema: BASE_MOTIONS (25) × STYLES (20) × SUBJECTS (20) random combinations

Intended use

• Training VLA / motion-tracking / imitation policies on the G1
• Pretraining kinematic priors for BeyondMimic-style RL loops
• Benchmarking text-to-motion → physics tracking pipelines
• Cross-embodiment motion transfer research

Limitations

• Motion is kinematic-first: physics tracking succeeds on ~99 % of episodes but a small tail of extreme prompts (rapid kicks, jumps) can produce mild foot-slip or root drift. No per-frame contact / balance labels included in v1.
• All episodes are 4.0 s (199 frames). Longer horizons will land in v2.
• No camera / RGB observations — proprioception + task text only.

Citation

@dataset{kimodo_g1_motion_v1_2026,
  author  = {Cagatay Cali},
  title   = {Kimodo → Unitree G1 Motion Dataset v1},
  year    = {2026},
  url     = {https://huggingface.co/datasets/cagataydev/kimodo-g1-motion-v1},
  note    = {10k text-conditioned humanoid episodes, LeRobot v3 format,
             generated via Kimodo diffusion + ProtoMotions GTP + MuJoCo}
}

Please also cite:

• Kimodo: nvidia/Kimodo-G1-RP-v1
• ProtoMotions / BeyondMimic: NVIDIA GEAR Lab
• LeRobot: Cadene et al., HuggingFace
• strands-labs/robots: agent-first robotics framework

License

Apache-2.0 for the dataset. Underlying models (Kimodo, ProtoMotions) retain their own licenses.