cagataydev/kimodo-g1-longhorizon-v3

Kimodo G1 Long-Horizon v3

10,000 episodes • 21.3M frames • 50Hz • 29-DoF Unitree G1

Text-conditioned whole-body humanoid trajectories generated with nvidia/Kimodo-G1-RP-v1 and tracked by cagataydev/protomotions-gtp-unitree-g1. Each episode chains ~50 primitive motions into a single ~3-minute "long-horizon day" of the humanoid living through 7 activity families (warmup → work → greet → patrol → combat → dance → cooldown → spontaneous).

Built with 🦆 strands-labs/robots — the humanoid robotics toolkit for language-driven agents.

---

What's in it

Field	Value
Episodes	10,000
Total frames	21,290,925
Mean frames/episode	2,129 (43s @ 50Hz)
Longest episodes	5,265 frames (~175s) — see mega_ep* samples
Robot	Unitree G1 (29 actuated DoF, 33 body links)
FPS	50
Format	LeRobot v3.0 (parquet chunks of 1000 episodes)
Codebase version	v3.0

Feature schema (meta/info.json)

{
  "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]},
  "task_index":        {"dtype": "int64",   "shape": [1]},
  "index":             {"dtype": "int64",   "shape": [1]}
}

• action — 29-DoF joint position targets for G1 PD controller (@50Hz outer loop, 1kHz inner). Joint order: left_leg(6) + right_leg(6) + waist_yaw/roll/pitch(3) + left_arm(7) + right_arm(7).
• observation.state — 65-dim proprioceptive state (root pose + linear/angular vel + 29 qpos + 29 qvel).

Family distribution (10K episodes)

Family	Episodes	Description
dance	1,468	celebration / rhythmic sway / twist / clap
workflow	1,453	pick / place / pull / push / carry
patrol	1,465	walk / run / turn / walk_backward / sidestep
social	1,400	wave / bow / salute / clap / meditate
warmup	1,403	stretch / sway / squats / arms_up
combat	1,396	boxing / kick / block / duck
mixed	1,415	randomized across families

Each episode composes ~50 primitive motions from the Kimodo motion library. Example task string:

MEGA long-horizon day: warmup with sway → jump → twist → sway → stretch → jump → reach forward; then work on tasks: reach forward → pull → pickup → crouch → pull → pickup → walk → pickup; greet colleagues: clap → arms up → salute → bow → arms up; patrol the perimeter: crouch → run → walk backward → jump → turn → walk backward → walk → crouch; combat drill: jump → sidestep → kick → sidestep → jump → sidestep → jump; celebration dance: sway → dance → jump → clap → arms up → sway → jump → sway; cooldown and rest: tpose → sway → meditate → sway; spontaneous moves: boxing → arms up → pull

---

Generation pipeline

text prompt (composed motion chain, ~50 primitives)
  → Kimodo diffusion (nvidia/Kimodo-G1-RP-v1, 100 steps)
  → per-primitive kinematic reference @ 30Hz
  → SLERP retime + stitch across primitives → 50Hz
  → ProtoMotions GTP ONNX tracker (cagataydev/protomotions-gtp-unitree-g1)
      • BeyondMimic-style RL tracker (arxiv:2408.07295)
      • Feed: current qpos/qvel + 4-step lookahead of mimic reference
      • Emit: 29-DoF PD targets + stiffness + damping
  → MuJoCo physics @ 1kHz (decimation=20 → 50Hz control)
  → LeRobot v3 parquet writer

Hardware: NVIDIA Thor (Jetson AGX, aarch64). Wall time ~1.5s / primitive motion; ~90s per full 50-primitive episode.

---

Reproducing episodes with 🦆 strands-robots

The full pipeline is available as a native provider in strands-robots via KimodoPolicy + ProtoMotionsPolicy (whole-body tracker). Kimodo is a whole-body 29-DoF kinematic generator — do NOT compose with WBC (which would override legs+waist). Use the matched ProtoMotions GTP tracker.

Installation

pip install "strands-robots[sim-mujoco,lerobot,kimodo,protomotions]"
export STRANDS_TRUST_REMOTE_CODE=1  # required — Kimodo uses HF trust_remote_code
export MUJOCO_GL=egl                # headless rendering (Jetson/servers)

Generate a single long-horizon episode

from strands_robots import Robot, create_policy

sim = Robot("g1", mesh=False)  # MuJoCo G1 (29-DoF, 33 bodies)

# Head-mounted camera on torso_link (G1 has no head_link — use offset)
sim.add_camera(
    name="head",
    parent_body="g1/torso_link",
    position=[0.08, 0.0, 0.35],
    target=[1.5, 0.0, 0.15],
    fov=70.0, width=640, height=480,
)
sim.add_camera(name="front", position=[3.0, 0.0, 1.2], target=[0.0, 0.0, 0.8])

sim.start_recording(
    repo_id="my-org/my-kimodo-longhorizon",
    root="./dataset",
    fps=50,
    cameras=["front", "head"],
    task="long-horizon humanoid day",
)

sim.run_policy(
    robot_name="g1",
    policy_provider="kimodo",
    policy_config={
        "model_id": "nvidia/Kimodo-G1-RP-v1",
        "diffusion_steps": 100,
        "num_frames": 120,        # per-primitive kinematic ref
        "native_fps": 30,
        "tracker_fps": 50,
        "device": "cuda",
    },
    tracker_provider="protomotions",
    tracker_config={
        "onnx_path": "hf://cagataydev/protomotions-gtp-unitree-g1/unified_pipeline.onnx",
        "yaml_path": "hf://cagataydev/protomotions-gtp-unitree-g1/unified_pipeline.yaml",
    },
    instruction=(
        "MEGA long-horizon day: warmup with sway → jump → twist → stretch; "
        "work: pull → pickup → walk → pickup; greet: clap → arms up → salute → bow; "
        "patrol: crouch → run → walk backward → turn; combat: jump → sidestep → kick; "
        "dance: sway → clap → arms up; cooldown: tpose → meditate"
    ),
    n_steps=8000, control_frequency=50, n_episodes=1,
)

sim.stop_recording()

Stream 10K episodes to HF Hub (this dataset's exact recipe)

from strands_robots.datasets import StreamingLeRobotWriter
from strands_robots import Robot

writer = StreamingLeRobotWriter(
    repo_id="cagataydev/kimodo-g1-longhorizon-v3",
    fps=50, chunks_size=1000, private=False,
    push_every_n_episodes=500,   # shard-push to HF
)

for ep_idx in range(10_000):
    task = compose_longhorizon_task(seed=ep_idx)  # ~50 primitives
    sim = Robot("g1", mesh=False)
    sim.run_policy(
        robot_name="g1",
        policy_provider="kimodo",
        tracker_provider="protomotions",
        instruction=task, n_episodes=1, control_frequency=50,
        writer=writer,
    )

See huggingface.co/blog/amazon/strands-lerobot-streaming-data-loop for the streaming-loop pattern (shard-push, resumable, distributed).

---

Training with this dataset

from strands_robots.training import create_trainer, TrainSpec

trainer = create_trainer(
    "lerobot_local",
    policy_type="pi0",  # or "smolvla", "diffusion", "act"
)

spec = TrainSpec(
    dataset_repo_id="cagataydev/kimodo-g1-longhorizon-v3",
    output_dir="./ckpts/pi0_g1_longhorizon",
    steps=200_000,
    batch_size=64,
    fps=50,
    method="full",
    extra={
        # For SmolVLA / Pi0.5 to unfreeze vision head — use Python booleans!
        "policy.freeze_vision_encoder": False,
        "policy.train_expert_only":     False,
    },
)
trainer.train(spec)

Verify num_learnable_params in the first log line:

• ~100M → expert-only, vision frozen
• ~393M → vision-unfrozen ✓

Pass Python False (not the string "false") — string values are truthy and silently keep the encoder frozen.

---

Sample episodes

Three long-horizon (50-primitive) sample episodes are shipped as MP4s + qpos.npz + frames.npy in this dataset's main_run/ directory:

• mega_ep0.mp4 (21.6 MB, 5265 frames, 175s) — sway → jump → twist → ... → arms_up → pull
• mega_ep1.mp4 (20.4 MB, 5265 frames) — sway → stretch → squats → ... → bow → crouch
• mega_ep2.mp4 (21.6 MB, 5265 frames) — arms_up → jump → squats → ... → run → boxing

Each MP4 shows the tracked G1 in physics doing the full ~50-primitive sequence in one continuous take, no cuts.

---

Known contracts

1. Whole-body vs arm-only generators: Kimodo emits 29 DoF, so it drives everything. Composing with WBC (velocity-conditioned walker) would override legs+waist, discarding 15/29 of the motion. ProtoMotions GTP is the matched tracker.
2. Verify parquet truth — never trust status=OK alone. Always run:

   import json, pyarrow.parquet as pq
   info = json.load(open("meta/info.json"))
   assert info["total_episodes"] == 10_000
   ep_tab = pq.read_table("meta/episodes/chunk-000/file-000.parquet")
   assert ep_tab.num_rows > 0
   

3. G1 URDF: 29-DoF variant (waist yaw+roll+pitch, wrists with roll+pitch+yaw, no fingers — rubber hands). g1_29dof_rev_1_0 URDF.
4. Anchor body for GTP tracker: torso_link (idx 16 of 33). Root: pelvis (idx 0).

---

Related

• 🤖 Robot policy tracker: cagataydev/protomotions-gtp-unitree-g1
• 🌀 Motion generator: nvidia/Kimodo-G1-RP-v1
• 🦆 Framework: github.com/strands-labs/robots
• 📖 Streaming datasets guide: huggingface.co/blog/amazon/strands-lerobot-streaming-data-loop

---

Citation

@dataset{cagataycali_kimodo_g1_longhorizon_v3_2026,
  author       = {Çağatay Çalı},
  title        = {Kimodo G1 Long-Horizon v3: 10K text-conditioned humanoid trajectories},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/cagataydev/kimodo-g1-longhorizon-v3},
  note         = {Generated with strands-labs/robots + nvidia/Kimodo-G1-RP-v1 + cagataydev/protomotions-gtp-unitree-g1}
}

License

Apache-2.0 — trajectories are generated by Kimodo (Apache-2.0) and tracked by our ProtoMotions GTP checkpoint (Apache-2.0). Free to use for research and commercial.