cagataydev/kimodo-g1-longhorizon-v2

Kimodo G1 Long-Horizon v2

10,000 episodes · 9.74M frames · 65 hours of Unitree G1 motion · text-conditioned, long-horizon

The largest open text-conditioned long-horizon locomotion+manipulation dataset for the Unitree G1 humanoid. Each episode is a chained multi-skill sequence (3–8 atomic motions) generated end-to-end from a natural-language prompt via a Kimodo → ProtoMotions GTP → MuJoCo pipeline, then recorded in LeRobot v3.0 format at 50 Hz.

Companion to cagataydev/kimodo-g1-longhorizon-v1 — v2 is the full 10k episode set with balanced family sampling and richer state features (65-dim vs v1's 29).

---

TL;DR

Robot	Unitree G1 (33 bodies, 29 DoF)
Total episodes	10,000
Total frames	9,739,525
Total unique tasks	9,190 (text prompts)
FPS	50
Duration	~54.1 hours of motion
Episode length	549–1424 frames (median ~899, ~18s @ 50Hz)
Action space	29-dim (whole-body PD targets)
State space	65-dim (root pose + 29 qpos + 29 qvel + gyro)
License	Apache-2.0
Format	LeRobot v3.0 (parquet)

---

What's inside — 7 task families

Each episode's prompt is composed by stringing together atomic motion slugs into a long-horizon sequence. Families are balanced to ~1400 episodes each:

Family	Count	Description	Example prompt
dance	1,468	Rhythmic / expressive sequences	"perform a dance sequence: sway → twist → dance → clap → dance → twist"
workflow	1,453	Task-oriented action chains	"workflow task: reach forward → push → pickup → walk → pull"
social	1,400	Greetings, salutes, gestures	"social greeting sequence: meditate → bow → arms up → wave"
combat	1,396	Fighting drills, boxing, kicks	"combat drill: kick → boxing → jump → sidestep → crouch → push"
patrol	1,465	Locomotion patrols with sidesteps	"patrol the area: walk → walk backward → sidestep → jump → crouch"
warmup	1,403	Stretches, squats, twists	"warmup routine: stretch → squats → twist → arms up → sway"
mixed	1,415	Deliberately cross-family chains	"long horizon mixed activity: tpose → meditate → bow → walk backward → salute → stretch → clap"

Atomic slugs (28 total) the composer draws from: walk, walk_backward, run, jump, sidestep, turn, crouch, still, wave, bow, salute, clap, arms_up, reach_forward, push, pull, pickup, kick, boxing, dance, sway, twist, stretch, squats, tpose, meditate, sit, stand.

---

The pipeline (text → physics)

prompt
  └─▶ Kimodo diffusion (nvidia/Kimodo-G1-RP-v1, 100 steps @ ~11 it/s)
       └─▶ 120 kinematic frames @ 30Hz   (root_pos + 29-dof qpos)
            └─▶ SLERP retime → 199 frames @ 50Hz
                 └─▶ ProtoMotions GTP ONNX tracker @ 50Hz
                      (cagataydev/protomotions-gtp-unitree-g1, BeyondMimic-trained)
                      └─▶ Unitree G1 MuJoCo physics @ 1kHz (decimation=20)
                           └─▶ LeRobot v3.0 episode (parquet)

Key: Kimodo produces whole-body 29-DoF kinematic reference (legs + waist + arms). The GTP tracker consumes all of it and emits PD targets. No composite / WBC leg-override — that's the wrong pattern for whole-body generators. See pipeline notes below.

---

Dataset schema (LeRobot v3.0)

{
  "codebase_version": "v3.0",
  "robot_type": "unitree_g1",
  "total_episodes": 10000,
  "total_frames": 9739525,
  "total_tasks": 9190,
  "fps": 50,
  "features": {
    "observation.state": {"dtype": "float32", "shape": [65]},
    "action":            {"dtype": "float32", "shape": [29]},
    "timestamp":         {"dtype": "float32", "shape": [1]},
    "frame_index":       {"dtype": "int64",   "shape": [1]},
    "episode_index":     {"dtype": "int64",   "shape": [1]},
    "index":             {"dtype": "int64",   "shape": [1]},
    "task_index":        {"dtype": "int64",   "shape": [1]}
  }
}

Field semantics

• action (29,) — GTP-emitted joint position targets consumed by G1's built-in PD controller at 50Hz. Joint order: left_leg(6) + right_leg(6) + waist(3, yaw/roll/pitch) + left_arm(7) + right_arm(7).
• observation.state (65,) — Compact G1 proprioception:
  • root_pos(3) + root_quat_wxyz(4) = 7
  • qpos(29) — actuated joint positions
  • qvel(29) — actuated joint velocities
• task_index — Index into meta/tasks.parquet (9,190 unique text prompts).

Loading

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("cagataydev/kimodo-g1-longhorizon-v2")
print(len(ds), "frames across", ds.num_episodes, "episodes")

sample = ds[0]
# sample["action"]              → torch.Tensor(29,)  PD targets
# sample["observation.state"]   → torch.Tensor(65,)  root+q+qv
# sample["task_index"]          → task id
# ds.meta.tasks[sample["task_index"].item()] → text prompt string

Or stream directly:

from datasets import load_dataset

ds = load_dataset("cagataydev/kimodo-g1-longhorizon-v2", split="train", streaming=True)
for row in ds:
    action = row["action"]
    state  = row["observation.state"]
    ...

---

Reproducibility

Generation script

The end-to-end generation pipeline uses composer.py (family templates), stitcher.py (Kimodo → GTP → MuJoCo), and run_longhorizon.py (episode loop with HF streaming push) built on top of strands-robots.

Hardware & runtime

Recorded on NVIDIA Thor (aarch64, 128GB unified) at ~3.2× realtime in the physics step. Wall-clock: ~13 hours for the full 10k with checkpointing every 100 episodes and periodic Telegram progress pings.

Model dependencies

Component	Model / Source
Text-to-motion diffusion	nvidia/Kimodo-G1-RP-v1
Motion tracker	cagataydev/protomotions-gtp-unitree-g1
Physics	MuJoCo 3.x, G1 g1_29dof_rev_1_0 URDF
Recorder	LeRobot 0.6+ (v3.0 dataset format)

---

Pipeline notes

• Kimodo is whole-body. Do not compose it with a WBC balance policy — that overrides legs+waist and throws away 15/29 DoF. Match Kimodo with a whole-body tracker (GTP) trained on the same DoF layout.
• GTP tracker inputs (8): current_anchor_rot(4), current_dof_pos(29), current_dof_vel(29), current_root_local_ang_vel(3), historical_processed_actions(1,29), mimic_future_{anchor_rot(4,4), dof_pos(4,29), dof_vel(4,29)}. Lookahead steps [1, 2, 4, 8] control-ticks = [20, 40, 80, 160] ms at 50Hz.
• GTP outputs: joint_pos_targets(29) fed to G1's built-in PD (stiffness/damping also emitted per-tick).
• Anchor body: torso_link (idx 16 of 33). Root: pelvis (idx 0).
• Recording pattern: episode boundary is dataset_recorder.save_episode() on the LeRobot recorder — not sim.save_episode() (which doesn't exist). Verified via parquet truth on every 100-ep checkpoint.

---

Verification contract

Every dataset assertion in this README was verified against the pushed parquet, not LLM narration:

import json, pyarrow.parquet as pq
from huggingface_hub import hf_hub_download

info = json.load(open(hf_hub_download(
    "cagataydev/kimodo-g1-longhorizon-v2", "meta/info.json", repo_type="dataset")))

assert info["total_episodes"] == 10000
assert info["total_frames"]   == 9_739_525
assert info["fps"]            == 50
assert info["robot_type"]     == "unitree_g1"

---

Suggested uses

• Text-conditioned VLA / VLM training — grounded action targets for open-vocab humanoid instruction following.
• Motion prior / diffusion-policy pretraining — long-horizon coverage across 7 skill families.
• Behavior cloning + RL fine-tuning — 65-dim state matches Isaac Lab / MuJoCo humanoid observation conventions.
• Retargeting benchmarks — GTP anchor/root convention is common enough to retarget to G1-Edu, H1, or other humanoids.

---

Citation

If this dataset helps your work, please cite:

@dataset{kimodo_g1_longhorizon_v2_2026,
  title  = {Kimodo G1 Long-Horizon v2: 10k text-conditioned humanoid episodes},
  author = {Cagatay, Cali},
  year   = {2026},
  publisher = {Hugging Face},
  url    = {https://huggingface.co/datasets/cagataydev/kimodo-g1-longhorizon-v2}
}

Underlying models:

• Kimodo — NVIDIA GEAR, text-to-motion diffusion for G1 (Apache-2.0)
• ProtoMotions GTP — NVIDIA GEAR, BeyondMimic-trained motion trackers (arXiv:2408.07295)
• LeRobot — Hugging Face

---

Changelog

• v2 (this dataset) — 10,000 episodes, 9.74M frames, 65-dim state, 7 balanced families, LeRobot v3.0.
• v1 — Initial release, 29-dim state, single-family bias. Deprecated in favor of v2.

---

License

Apache-2.0. Data is derived from open-source models (Kimodo, ProtoMotions GTP) and simulated on the publicly available Unitree G1 URDF. Commercial and research use both permitted.