Hugging Face
The AI community building the future
The platform where the machine learning community collaborates
on models, datasets and applications (and hardware !)
Hugging Face logo
Hugging Face
The AI community building the future
🗄️ Datasets
Curated datasets for every ML taskSaves time in data collection and pre-processingBuilt-in viewer to explore the data
🧊 Models
Pre-trained ML modelsText, audio, image, video… and robot actionsDozens of libraries, including Transformers and LeRobot
🧩 Spaces
Build, host and share ML appsThousands of ML-powered demosGradio, Streamlit, Docker or static HTML
🪣 Buckets
Storage Buckets for big dataCheckpoints, logs and terabyte-scale datasetsLeRobot v0.7.0 trains straight from them
LeRobot
Bringing AI into the real world
Making AI for robotics more accessible
with end-to-end learning
LeRobot logo
LeRobot today
An open-source home for Physical AI
~300 TBof open Physical AI data on the Hub, in LeRobot format
~30kGitHub stars, and counting
20k+members on Discord and X
v0.7.0From terabytes to prompting robots
LeRobot: an end-to-end Physical AI workflow
1. Configure robots
Unified Python API
Open hardware designs, plug & play
YAM, reBot B601, Unitree G1 + SONIC
2. Teach & teleoperate
Leader-follower, VR, keyboard, phone
Record demonstrations in real time
…in simulation too, same commands
3. Build & stream datasets
Standardized schema, shared on the Hub
Language annotations (subtasks, VQA…)
Stream from Hub, Buckets or LanceDB
4. Train policies
ACT, Diffusion, VLAs (π0.5, GR00T, SmolVLA…)
World action models + 4 new VLAs
FSDP2 / HSDP, EMA, Trackio
5. Deploy
lerobot-rollout on the real robot
Talk to it: /subtask, /vqa, /autosteer
Remote inference on a GPU server
6. Evaluate & improve
Simulation benchmarks (LIBERO, …)
One sim core + RoboLab, Wilson error bars
Reward models to score your data
LeRobot v0.7.0
From terabytes to prompting robots
Every policy starts as terabytes of demos and ends as a robot following instructions

💬 Language in the loop

Switch subtasks, ask questions and autosteer a robot while it works

💾 Data at scale

Train on the Hub, Storage Buckets or LanceDB without downloading

⚡ Training at scale

FSDP2, HSDP, distributed checkpoints, FP16, EMA, Trackio

🌍 New models

4 world action models + 4 VLAs: Cosmos3, FLUX 3 Action, G0.5, LingBot-VLA 2.0…

🛰️ Remote inference

Big model on a GPU, light computer on the robot: lerobot-policy-server

🧪 Sim & hardware

Every simulator on one core · YAM, reBot B601, G1 + SONIC

1 - Robots
Real robots…
SO-101 Arm
LeKiwi mobile base
New in v0.7.0
1 - Robots
More arms, more humanoids
I2RT

YAM

  • One arm or two: bimanual setups composed from single arms
  • Gravity feed-forward from the arm's inertial model
Seeed Studio

reBot B601

  • 6-DoF arm, paired with the reBot Arm 102 leader
  • RobStride or Damiao motors with one motor_family setting, even mixed bimanual
Unitree × NVIDIA

G1 + SONIC

  • Policy outputs a 64-d motion token instead of 29 joint targets
  • Onboard decoder → 50 Hz whole-body joint targets
  • Works with remote inference: policy on a GPU, decoder on the G1's Jetson
  • New 23-DoF G1 variant (GR00T & Holosoma locomotion)
2 - Control & teleoperation
3 - Datasets
The ChatGPT moment of robotics?
Cumulative number of LeRobot datasets on the Hub over time
~300 TB
of open Physical AI data in LeRobot format
New in v0.7.0
3 - Datasets
Train on terabytes without downloading them

Disk-less streaming

2.1–2.5×samples/s with 4 DataLoader workers

Fetch only the video byte ranges you need, from the Hub or a Storage Bucket. Every frame once per epoch, deterministic resume.

LanceDB backend

local-disk speedfrom object storage, true global shuffle

Lance tables keep the original MP4s as blobs: nothing re-encoded, training starts right away.

GPU augmentation

2.8× / 3.3×faster ACT training on H100 / A100

Augment once per batch on the GPU, plus 8 robotics-oriented transforms (motion blur, shadows, JPEG…).

from lerobot.datasets import LeRobotDataset
dataset = LeRobotDataset("lance-format/pusht-lance")  # same API, Lance under the hood
# or: lerobot-train --dataset.streaming=true --dataset.image_transforms.backend=gpu ...
New on the Hub
3 - Datasets
Some of the largest open robot datasets landed in the past months
WANDA
LeCAR Lab (CMU)
Synthetic demos from 1 real demo/task
~248 h0.27 TB
AXIS Franka v2
AXIS Robotics
Sim, browser teleop, 1,000+ tasks
~3,300 h11.0 TB
OpenNeoData
NeoteAI
Visuotactile, 5 platforms + handhelds
5,000+ h9.1 TB
HiFi-UMI-2K
Simple World Lab
Robot-free capture, 6 camera views
2,000 h16.0 TB
PrimeBot Household
PrimeBot / Swancor
Laundry & clothes folding, bimanual
700+ h5.2 TB
EgoSuite-Open100K
LightWheel
Egocentric human video + 3D pose
100,000 h13+ TB
EgoSteer-RealWorld
EgoSteer
Glove teleop in RGB-D, dexterous hands
192 h3.2 TB
RW-RL-HIL
BodenAI
Deployments with human takeovers
82 h0.19 TB
EgoSteer-Egocentric
EgoSteer
3D hand labels for 8 ego datasets
~3,200 h0.40 TB
SharpaDex v1.0
Sharpa Robotics
Tactile + force, 22-DoF hands
300 h7.7 TB
RoboQuest
DeCLaRe Lab
Sim, explore-then-act kitchen tasks
366 h0.34 TB
FineART
Scale AI × Hugging Face
530k+ human-written subtask labels
1,718 h1.9 TB
4 - Models
A first step in AI models for robotics
ACT architecture diagram
Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware
Action Chunking with Transformers (ACT)
4 - Models
Vision-Language-Action models (VLAs)
GR00T N1.7 architecture diagram
NVIDIA Isaac GR00T N1.7 - A Foundation Model for Generalist Robots
4 - Models
Policy rollouts from the community
New in v0.7.0
4 - Models
The VLA zoo keeps growing
OpenGalaxea

G0.5

Qwen3.5-2B VLA that thinks out loud: chain of thought (subtask, boxes) and actions in one stream. System 1 or System 2.

Non-commercial license

Robbyant

LingBot-VLA 2.0

Qwen3-VL-4B + sparse MoE action expert (32 experts, top-4), 55-d cross-embodiment actions.

756 → 120 msper chunk with CUDA graphs + compile
Dexmal

DM0.5

Gemma3-4B VLM + 680M flow-matching action expert, ported from OpenDM onto standard LeRobot.

197/200LIBERO reproduction check
X Square Robot

Wall-OSS-0.5

4B VLA, now native on Transformers. Its VL expert generates text: answers /vqa and plans its own subtasks.

New in v0.7.0
4 - Models
World action models: imagine, then act
Predict the future (video or latents) and the next action chunk, together
NVIDIA

Cosmos3

Cosmos 3 world foundation models for robots. Nano (16.5B) and Edge (4.6B): full FSDP2 or LoRA fine-tuning.

97.9%LIBERO avg. (Nano & Edge)
Black Forest Labs

FLUX 3 Action

Frozen video VAE + text encoder feed a diffusion transformer; only actions come out at inference.

base · SO-101 · DROIDcheckpoints, inference + LoRA
RLinf

LaWAM

Dynamics in latent space: Qwen3-VL backbone, latent action & world model, flow-matching head.

98.5%LIBERO (1,970 / 2,000)
NVIDIA LongLive

Long-WAM

Built for long context. IDM imagines future latents without decoding a frame; COD denoises video + actions jointly.

99.0%LIBERO avg. (IDM)
New in v0.7.0
4 - Models
A training stack built for 6B–16B parameter models

FSDP2, HSDP & distributed checkpoints

Sharding, gradient accumulation, PEFT under sharding, and checkpoints that resume on a different number of GPUs.

3.81×faster checkpoint saves
42.6 GBpeak/GPU on resume (was OOM at 143.8 GB)

6B FastWAM on 8×H200

Large-model recipes

  • FP16 with loss scaling on every topology, one dtype for all policies
  • EMA of policy weights
  • Warmup-stable-decay schedule
  • Fused Adam / AdamW

Trackio

Hugging Face's open-source experiment tracker replaces Weights & Biases.

  • No account needed locally
  • Syncs to a Space to share
  • Records launch command and console output, tracebacks included
New in v0.7.0
Talk to your robot
Language in the loop: interactive rollouts
$ lerobot-rollout --interactive=true ...
/start
/subtask pick up the red cube
/vqa how many cubes are left?
  → "Two: one blue, one green."
/autosteer clear the table
/stop
/subtask
Switch the instruction on the fly, without stopping
/vqa
Ask the policy what it sees
/autosteer
Hand it a goal, it picks its own subtasks until you take back control
Robot, weights and processors stay loaded between instructions. You can record episodes while you steer.
New in v0.7.0
Talk to your robot
Who does the thinking?
Scale AI × Hugging Face

FineART-VLA

π0.5 + a language head, trained on FineART: 40k+ bimanual episodes, 1.7k+ hours, 530k+ subtask annotations.

System 2

Predicts its next subtask in words

System 1

Flow-matching action expert carries it out

Knowledge insulation keeps action training from degrading the VLM.

External VLM planner

Any VLM behind an OpenAI-compatible API: transformers serve, vLLM or Inference Providers.

Turns "tidy up" into precise instructions your VLA understands, optionally restricted to the subtasks it was trained on.

Train your own

YAML recipes turn lerobot-annotate labels (subtasks, plans, memory, VQA, speech) into chat-style training targets.

Recipes ship inside the policy config, so training and inference render language the same way. Annotation runs on HF Jobs.

5 - Inference
Bimanual SO arms running a policy
New in v0.7.0
5 - Inference
Big models, light robots: remote inference

🤖 Robot computer

  • Laptop or single-board computer
  • Cameras, motors, teleop, recording
  • Executes buffered actions
lerobot-rollout --inference.type=remote ...
observations + instructions →
Zenoh · TLS · ACL
← action chunks

🖥️ GPU server

  • Runs the big VLA / WAM
  • Computes the next chunk while the robot moves
  • Real-time chunking (RTC) when supported
lerobot-policy-server --config_path=server.yaml
Overlapping inference and execution hides latency · expired steps dropped, overlaps blended · interactive sessions work remotely too
6 - Benchmark
New in v0.7.0
One simulation core
Every simulator behind one vector-env core: batched lerobot-eval, or a sim robot for rollout, recording and teleop. In-process or in Docker via lerobot-env-server.
RoboLab
NVIDIA's 120 tabletop tasks for DROID Franka on Isaac Lab
Error bars
Success rate 90.0% (9/10 episodes,
95% Wilson interval 59.6% to 98.2%)
What makes LeRobot special?
Its community, of course!
20k+Discord & X members
~30kGitHub stars

LeRobot LIVE

First edition on Discord on Oct 8. Next one in December!

What makes LeRobot special?
Its endless possibilities
Who is using LeRobot?
Logos of organizations using LeRobot
How to get started?
# Beginner mode
pip install lerobot

# Advanced mode
git clone https://github.com/huggingface/lerobot.git
cd lerobot
pip install -e ".[many-cool-options]"

# New in v0.7.0: talk to your robot
lerobot-rollout --interactive=true ...
And let's go 🚀
More than just LeRobot
Reachy Mini and a duck-shaped biped robot
LeLab@huggingface/lerobot for JavaScriptMachine-translated docsThird-party hardware & integrations
Thank you
LeRobot logo
github.com/huggingface/lerobot · huggingface.co/lerobot
← → navigate · F fullscreen · N notes