GR00T Fine-Tuning Data for the Unitree G1
GR00T fine-tuning data recorded by teleoperating the Unitree G1 at a table or walking, with the Dex1 gripper, tactile hands or the Dex3, Inspire and BrainCo Revo 2 hands. Each episode pairs camera views, a task description, joint states and actions on one clock, delivered in LeRobot format for post-training GR00T N1 models.
What GR00T N1 Models Expect From Post-Training Data
GR00T N1.7, the newest GR00T N1 release, is an open vision-language-action model that NVIDIA describes as adaptable through post-training for specific embodiments and tasks. Its model card lists four inputs, and GR00T fine-tuning data supplies each one.
Camera frames
Frames from the robot's cameras, one MP4 stream per camera named observation.images.<view>.
Language instruction
A task instruction, listed in meta/tasks.jsonl and referenced by an annotation column.
Proprioceptive state
The joint state at each step, stored as observation.state.
Embodiment ID
An integer naming the embodiment. A custom setup registers as NEW_EMBODIMENT.
A flow-matching diffusion transformer outputs a chunk of continuous action vectors sized to the robot's degrees of freedom, trained against the dataset's action column.
The GR00T LeRobot Layout
NVIDIA's Isaac-GR00T repo says GR00T uses a flavor of the LeRobot v2 format plus meta/modality.json, which splits the state and action arrays into named fields and maps each video key to its camera. For the G1 the repo names UNITREE_G1, used with a decoupled whole-body controller, and UNITREE_G1_SONIC, the tag its end-to-end collect, fine-tune and deploy workflow supports.
How Many Demonstrations NVIDIA's G1 Workflow Asks For
NVIDIA's data collection guide for its GR00T and Unitree G1 workflow sets a clear bar for its example task, moving an apple onto a plate with the left hand.
At least 200 clean successes
The guide asks for at least 200 clean successful demonstrations collected manually. More high-quality runs broaden the set.
Around 200–400 timesteps
Each demonstration aims for around 200–400 timesteps. Overly long runs slow later pipeline stages and overly short ones often turn jerky.
Clean and consistent motion
Saved runs avoid unnecessary collisions and move steadily. The baseline set keeps objects, lighting and cameras fixed while broader real-robot sets vary object placement.
The Isaac-GR00T repo gives no general minimum. We size each G1 pilot around the tasks and variations your evaluation covers, with a success condition and reject rules written per task before recording.
What NeuralChainAI Records on the G1
We teleoperate the Unitree G1 with the end effector your task needs, so the state and action fields match the robot your GR00T policy runs on.
Tabletop and stationary work
The G1 holds its stance at a counter, table or shelf while its arms, waist and hands work.
Walking tasks
Loco-manipulation episodes record walking and manipulation together, adding leg joints to the arrays.
Dex1 gripper
A gripper for pick-and-place, packing and tool handling, which keeps the hand action space small.
Dex3, Inspire and BrainCo Revo 2 hands
Multi-finger hands for in-hand grasps, each actuated finger joint adding a state and action value.
Contact-sensing hands
Tactile hands add contact readings to each episode, with force-torque streams where the program calls for them.
Streams agreed per program
Camera views, depth, joints, actions, tactile and force-torque, chosen with you.
Learn more →GR00T Inputs and What Each NeuralChainAI Episode Provides
Each row pairs a GR00T N1.7 input, as NVIDIA documents it, with what a G1 episode carries.
| GR00T input | What NVIDIA's docs specify | What each G1 episode provides |
|---|---|---|
| Camera frames | A variable number of frames from robot cameras, one MP4 per camera | Head and left and right wrist views on one clock, with depth where included |
| Language instruction | A text string listed in meta/tasks.jsonl | A plain-language task description for every episode |
| Proprioceptive state | observation.state, split into named fields by meta/modality.json | Arm, waist and hand joint positions at every step, plus the legs in walking episodes |
| Actions | Continuous vectors sized to the robot's degrees of freedom | The commands sent at each step for the same joints |
| Embodiment | An integer ID, with NEW_EMBODIMENT for a custom setup | The G1 configuration and end effector, recorded with every episode |
| Dataset format | GR00T's flavor of LeRobot v2 plus meta/modality.json, with v3 converted to v2 | LeRobot format, with MCAP, HDF5 and RLDS on request |
Ready-Made G1 Sets and Custom Collection
Our ready-made G1 sets cover kitchen, bedroom and staged retail tasks, the retail tasks recorded with real teleoperation in a mock store set, so a GR00T team can post-train on real G1 episodes first. How the ready-made terms work:
A new task on the G1, or a new robot.
Custom collection covers a new task on a robot we already record with, or a new robot or embodiment, on non-exclusive or exclusive terms. Exclusivity covers the data and its resale rights, separate from the robot that recorded it. Exclusive terms are priced above non-exclusive, and term and scope are agreed once a pilot is defined.
Delivery in LeRobot Format for GR00T Post-Training
Episodes ship in LeRobot format, with MCAP, HDF5 and RLDS on request. Unitree's own unitree_IL_lerobot repository converts G1 recordings made with Dex1, Dex3, BrainCo and Inspire end effectors to LeRobot format and lists GR00T among the policies it trains.
LeRobot by default
Camera streams, joint states, actions and task descriptions in LeRobot format.
GR00T's modality file
NVIDIA's data preparation guide has you add meta/modality.json to a LeRobot v2 dataset, and its fine-tuning script reads a modality config naming your camera, state and action keys.
Provenance records
Every episode carries provenance and consent records you can audit.
For help with the post-training run, our physical AI consulting and development team fine-tunes GR00T and pi0 on real demonstrations, retargeted onto your embodiment. Our pages on humanoid robot training data, ACT and Diffusion Policy training data and training data for robot policies and VLAs cover other setups and policies.
What Shapes the Engagement
We scope every GR00T data program against these factors with you, then quote it as a whole.
Hands and end effector
The Dex1 gripper, a tactile hand or a Dex3, Inspire or BrainCo Revo 2 hand.
Tabletop or walking
Stationary work at a counter or table, or walking between stations with the legs in the episode.
Tasks and environments
The kitchen, bedroom and staged retail sets, or tasks written for your evaluation.
Camera views and streams
Which views and streams each episode carries, matched to your modality config.
Acceptance and QC criteria
What counts as a reject, such as a dropped object, a collision or a stream out of sync.
Pilot size
How many episodes and task variations the pilot covers for your team to evaluate.
Exclusivity and sublicense rights
Non-exclusive, sublicensable ready-made sets, or custom collection on non-exclusive or exclusive terms.
Have a GR00T Post-Training Task to Record?
Send us the task, environment and hands your G1 policy needs. We'll scope a pilot with the streams, calibration and format you'd get at scale, so your team can check the episodes first.
Book a Strategy Session →Frequently Asked Questions on GR00T Fine-Tuning Data
NVIDIA's GR00T N1.7 model card lists camera frames, a language instruction, proprioceptive state and an embodiment ID as inputs, with continuous action vectors as output. Each episode needs synced camera video, a task description, joint states and the commanded actions. Our G1 episodes carry all of these.
NVIDIA's guide for its GR00T and Unitree G1 workflow asks for at least 200 clean successful demonstrations of around 200–400 timesteps for its example task. Your count depends on the tasks and variations you evaluate, so we size the pilot with you.
LeRobot format, with MCAP, HDF5 and RLDS on request. GR00T reads a flavor of LeRobot v2 with an added meta/modality.json file, documented in NVIDIA's data preparation guide.
The Dex1 gripper, tactile hands or the Dex3, Inspire and BrainCo Revo 2 hands, fitted to your objects and grasps. A gripper usually adds one open and close value per hand, while a multi-finger hand adds a value per actuated finger joint, so we agree the end effector before the pilot.
Ready-made G1 sets are licensed non-exclusive, and you can sublicense them to your own client. Custom collection runs on non-exclusive or exclusive terms, with exclusivity covering the data and its resale rights and priced above non-exclusive. Term and scope are agreed once a pilot is defined.
Related Physical AI Solutions
Humanoid Robot Training Data and Teleoperation
Learn more →Pi0 Fine-Tuning Data for openpi
Learn more →ACT and Diffusion Policy Training Data for Robots
Learn more →Embodied AI Training Data for Robot Policies
Learn more →Robot Data Collection Services
Learn more →Physical AI and Robotics Consulting and Development
Learn more →Physical AI and Robotics Consulting
Help with the GR00T post-training run on your embodiment.
See physical AI services →AI Strategy Session
A 30-minute scoping call on your task, hands, camera views and format.
Book a session →Humanoid Robot Training Data
Standing and walking G1 episodes with grippers or multi-finger and tactile hands.
See humanoid robot training data →Scope Your GR00T Fine-Tuning Data
A 30-minute strategy call. Tell us the task, the hands and the camera views your GR00T policy needs, and we'll scope a G1 pilot with the streams, format and acceptance criteria to post-train on.