explainx / blog
BitRobot, Unitree, and Hugging Face released HIW-500 — 500+ hours, 23K+ episodes, 10+ TB of Unitree G1 teleop in Southeast Asian homes. LeRobot format cuts it to 2.15 TB. Full guide.
Jul 14, 2026
Xiaomi's 38B autoregressive world foundation model keeps general T2I and editing in the training mix while learning multi-view robot scene synthesis. Synthetic embodied transfer data nearly doubles π₀.₅ out-of-distribution success on real manipulation tasks. explainx.ai breaks down tasks, architecture, and limits.
Jul 10, 2026
1X calls NEO's new hands "an API to the physical world" — 25 force-controlled DoF, quasi-direct-drive tendons, tactile shear sensing, and a 10K/year production line. explainx.ai breaks down the July 9 launch with hosted demo clips.
Jul 9, 2026
Mistral AI's first robotics model takes plain-language instructions and one front camera feed to navigate offices and warehouses—map-less, no depth sensors, 76.6% success on unseen R2R-CE benchmarks. It is not open yet, but the HN thread surfaced the real deployment questions.
On June 23, 2026, BitRobot (@BitRobotNetwork) released HIW-500 — Humanoids-in-the-Wild 500 — calling it the largest open-source humanoid teleoperation dataset collected in real homes.
Built with Unitree and Hugging Face / LeRobot, the drop lands the same day as Nori L2's iPhone-priced robot teaser and a week after Genesis Eno. The through-line in June 2026 robotics: real environments, not lab demos, are where the data wars are being fought.
"Built for learning from real homes, not lab-only scenes." — HIW-500 project page
TL;DR
| Spec | HIW-500 |
|---|---|
| Release | June 2026 (V1) |
| Robot | Unitree G1 (29-DoF + grippers) |
| Collection | 12 real homes, Southeast Asia |
| Hours | 500+ teleoperation |
| Episodes | 23,743 |
| Frames | 40,839,947 |
| Tasks | 11 household skills |
| Subtask labels | 161 labels, 148K+ annotations |
| Raw size | ~10+ TB (ROS bag / MCAP) |
| LeRobot size | ~2.15 TB (v3.0, same trajectories) |
| License framing | Research + commercial training (see HF page) |
| Download | BitRobot/HIW-500-LeRobot |
Lab datasets teach robots to repeat. Home datasets teach robots to generalize.
HIW-500 explicitly targets variation that controlled benchmarks strip out:
That is the gap between Figure Helix-02 tidying a staged bedroom and a robot that works in your kitchen. Policy researchers have argued for years that pooling diverse teleop data drives generalist policies — Open X-Embodiment showed cross-embodiment pooling improves small-data domains by 50%. HIW-500 is the humanoid-home slice of that bet at scale.
Unitree's response on X:
"We're excited to support BitRobot in open-sourcing the largest humanoid whole-body teleoperation dataset collected in real homes. We hope it accelerates progress toward general-purpose humanoid robots." — @UnitreeRobotics
Each episode records human whole-body teleoperation of a Unitree G1 in a real home — not joystick-only arm control, but locomotion + bimanual manipulation together.
| Stream | Resolution | FPS | Notes |
|---|---|---|---|
| Head camera | 480 × 1280 RGB | 30 | Stereo scene context, navigation |
| Left wrist | 480 × 640 | 30 | Stereo IR, close-range manipulation |
| Right wrist | 480 × 640 | 30 | Stereo IR, close-range manipulation |
Videos are encoded AV1 in the LeRobot release (yuv420p, CRF 30).
From the LeRobot dataset card:
observation.state — 29-DoF joint positions (hips, knees, ankles, waist, arms, wrists)observation.state.wbc — whole-body control state: pivot velocity, roll/pitch/yaw, height, left/right end-effector poses, gripper triggersaction — 23-D whole-body teleop commands mirroring human operator inputslanguage_persistent / language_events — language annotations per episodeThis is VLA-ready multimodal data: three video streams + proprioception + language labels in a single LeRobot v3.0 schema (codebase_version: v3.0, robot_type: unitree_g1).
BitRobot's project page lists every task in V1:
| Task | Category |
|---|---|
| Building children's table | Assembly / furniture |
| Hang hanger | Closet organization |
| Clean up the room | General tidying |
| Setting the table | Dining prep |
| Restocking fridge | Kitchen manipulation |
| Kitchen organization | Storage / sorting |
| Hang keys on a hook | Fine placement |
| Move pillow to sofa | Soft-object handling |
| Sweep floor | Tool use + navigation |
| Picking trash | Grasp + dispose |
| Clothes washing | Laundry workflow |
161 subtask labels and 148,000+ subtask annotations break episodes into fine-grained action tiles — useful for hierarchical policies, skill discovery, and evaluation rubrics rather than end-to-end black-box imitation only.
| Format | Dataset | Size | Best for |
|---|---|---|---|
| Raw | BitRobot/HIW-500 | ~10+ TB | Full ROS bag / MCAP, custom pipelines |
| LeRobot v3.0 | BitRobot/HIW-500-LeRobot | ~2.15 TB | ACT, diffusion policies, HF training stacks |
LeRobot's team re-encoded the full corpus:
"We re-encoded the full dataset into LeRobot format: ~10TB → ~2TB, no loss of fidelity. Same trajectories, a fraction of the footprint, far easier to stream." — @LeRobotHF
For researchers already on LeRobot — the same stack the Columbia Nori Bot paper uses for ACT training — HIW-500-LeRobot is the practical entry point. Parquet chunks at 100 MB, video chunks at 200 MB, 1,000 episodes per chunk.
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
dataset = LeRobotDataset("BitRobot/HIW-500-LeRobot")
print(dataset.num_episodes) # 23743
print(dataset.num_frames) # 40839947
sample = dataset[0]
print(sample.keys()) # observation.images.*, action, language_*, etc.
Verify against your installed LeRobot version — HIW-500 targets LeRobot v3.0.
| Partner | Role |
|---|---|
| BitRobot | Dataset lead; "world's open robotics lab" |
| Unitree | G1 hardware platform |
| Hugging Face / LeRobot | Hosting + LeRobot conversion |
| Virtuals Protocol | Public supporter on X |
BitRobot's roadmap shows V1 in June 2026 (this release) and V2 with more tasks and environments later.
| Layer | HIW-500 role |
|---|---|
| Imitation learning | 500+ hours of human demos → ACT / diffusion / VLA training |
| Mobile manipulation | Whole-body teleop — walk, reach, grasp in one trajectory |
| Generalist policies | Home diversity beats lab repetition for OOD generalization |
| Simulation bridge | Feed NVIDIA Cosmos or Genesis-style sim for sim-to-real |
| Data collection ops | Contrast with Shift-free cleaning data collection in NYC — different geography, same "real world" thesis |
Consumer vs research gap: Nori L2 targets sub-$1k manipulators with OpenClaw scheduling. HIW-500 targets Unitree G1-class humanoids learning from hundreds of home hours. Same month, opposite ends of the cost curve — both need data.
@misc{hiw500_2026,
title={HIW-500: Humanoids In-the-Wild Dataset for Robot Learning},
author={BitRobot and Unitree and Hugging Face},
year={2026},
howpublished={\url{https://bitrobot-foundation.github.io/humanoids-in-the-wild-500-hours/}}
}
Project page includes a Rerun live preview for browsing episodes before downloading terabytes.
For commercial access beyond the public release, BitRobot offers Request Data Access on their project site.
Humanoids and datasets:
Physical AI and training:
Primary sources: HIW-500 project page · LeRobot dataset · Raw dataset · @BitRobotNetwork · @UnitreeRobotics · @LeRobotHF
HIW-500 is the largest public humanoid teleop dataset from real homes as of June 23, 2026: 500+ hours, 23,743 episodes, 11 household tasks, 12 Southeast Asian homes, Unitree G1 whole-body control, and 148K+ subtask annotations. Download raw at ~10 TB or LeRobot v3.0 at ~2.15 TB on Hugging Face — same trajectories, easier streaming.
For anyone training mobile manipulation, bimanual skills, or home generalization policies, this is the dataset drop to benchmark against. Lab scores still matter; in-the-wild hours are what close the gap between demo videos and deployable humanoids.
Episode counts, file sizes, and task list reflect BitRobot's June 2026 V1 release and the Hugging Face dataset card. Re-check huggingface.co/datasets/BitRobot/HIW-500-LeRobot before large downloads — the corpus is multi-terabyte.