Unitree open-sources UnifoLM-WBT-Dataset — humanoid whole-body teleoperation dataset

TL;DR

Unitree open-sourced the UnifoLM-WBT-Dataset, a real-world humanoid whole-body-teleoperation dataset collected on the Unitree G1 platform, hosted on Hugging Face. Unlike conventional manipulation datasets that record arms separately from legs, WBT episodes capture the entire robot — bipedal locomotion, balance, and dexterous manipulation — as a coordinated system, driven by XR / MoCap / exoskeleton teleoperators. The collection went public on March 5, 2026 and Unitree promises high-frequency rolling updates, positioning it as a candidate “Open X-Embodiment for humanoids.” Paired with the existing UnifoLM-WMA-0 (world-model + action head) and UnifoLM-VLA-0 releases, it is the dataset half of an end-to-end open humanoid stack.

Key claims

The release is a real-world humanoid robot whole-body-teleoperation (WBT) dataset for open environments, opened on 2026-03-05 with promised high-frequency rolling updates [tweet body].
The stated aim is the most comprehensive real-world humanoid dataset in terms of scenario coverage, task complexity, and manipulation diversity [tweet body].
Hardware is the Unitree G1 humanoid (29–43 DoF depending on hand variant), driven by XR teleop (Apple Vision Pro, PICO, Meta Quest), MoCap, or wearable exoskeleton-style controllers [HF collection page / Warmcore press summary].
Each episode is RLDS-formatted with full joint-state observations and action vectors at 30 fps, plus synchronized multi-camera video at 256×256 or 128×128, episodes averaging ~30 s [HF dataset cards].
The collection is structured as task-specific sub-datasets (Dex1, Dex3, BrainCo, Z1 arm, etc.) rather than a single monolithic file [HF unitreerobotics org page].

Method

This is a product/dataset announcement, not a research paper, so there is no novel method to summarize. The relevant mechanics are the data-collection pipeline:

Operators wear XR headsets / MoCap suits / exoskeleton input devices.
Human kinematics are mapped directly into joint-space commands for the G1’s 29–43 DoF body.
The robot executes whole-body behavior in real environments while a synchronized recording captures (i) full joint observations, (ii) action vectors, (iii) multi-camera RGB at 30 fps, (iv) task labels.
Trajectories are written in RLDS — the same format used by Open X-Embodiment and LeRobot — so the data drops in to existing VLA training stacks without conversion.

The companion UnifoLM-WMA-0 repo positions this dataset as the training corpus for a world-model + action architecture, where the learned world model serves as both interactive simulator and policy enhancer.

Results

No numbers are claimed in the tweet. Adoption signal as of the tweet date: the HF collection lists eight sub-collections (UnifoLM_WBT_Dataset, UnifoLM_G1_Dex1_DiverseManip_Dataset, UnifoLM-VLA-0, UnifoLM-WMA-0, UnifoLM_G1_Dex1_Dataset, UnifoLM_G1_Brainco_Dataset, UnifoLM_Z1_Arm_Dataset, UnifoLM_G1_Dex3_Dataset) with sub-dataset viewer counts in the hundreds-of-thousands of rows. The tweet itself reported 5.6M views by the time of filing.

Why it’s interesting

For Luma’s embodied-world-model work, an open, RLDS-formatted, whole-body humanoid dataset is rare — most open robot datasets are arm-only (Open X-Embodiment, LeRobot pi0 corpora) and most whole-body data sits behind humanoid-company walls (Figure, 1X, Tesla Optimus). This complements Fauna Sprout: A lightweight, approachable, developer-ready humanoid robot, which delivered the hardware axis of a research-friendly humanoid stack but no released data; together with UnifoLM-WMA-0 it covers the dataset+world-model halves that Sprout punts on. It is also the missing ingredient under the VLA models candidate cluster (Isaac GR00T, OpenVLA, π0, Alpamayo) — those models can now train on a real-world bipedal corpus without licensing one bespoke. The rolling-update model is unusual and worth tracking: most robot datasets are frozen at release.