XR-1 (X Robotic Model 1)

XR-1 is the manipulation policy of the Beijing Innovation Center of Humanoid Robotics — the state-backed unit behind the Tiangong humanoids — released as open weights in December 2025 under the MIT licence. Its idea is a detour. Most vision-language-action models map camera images and an instruction straight onto joint commands. XR-1 inserts a shared vocabulary in between: Unified Vision-Motion Codes, a discrete codebook of 256 entries learned by a two-branch VQ-VAE that encodes what changes in the picture and what the arm does at the same time. Because that vocabulary is written in terms of dynamics rather than in terms of any one machine, the authors can pour human demonstrations and recordings from unlike robots into the same training run and have them reinforce each other instead of fighting. The model ships in two pieces and the catalogue records both: XR-1-Stage1-UVMC, the 1.18-billion-parameter tokeniser that produces the codes, and XR-1-Stage2, the 4.11-billion-parameter policy that consumes them. Stage 2 reads up to six camera views at 224 x 224, a 32-dimensional robot state and a short instruction of at most 48 tokens, and answers with a chunk of 50 future actions in ten denoising steps. Pre-training used XR-D, the centre's own cross-embodiment set of 158,000 trajectories and 69.1 million frames. The evaluation is unusually large for the field: over 14,000 real-world rollouts across 120-plus tasks on six different machines — the centre's own Tiangong 1.0 and 2.0, single- and dual-arm UR-5e, dual-arm Franka and an AgileX Cobot Magic 2.0 — with reported wins over pi-0.5, pi-0, RDT, UniVLA and GR00T-N1.5. The paper was accepted as an oral at ICML 2026. One editorial caveat about the paperwork rather than the model: both Hugging Face repositories point readers to arXiv 2411.02776, and that identifier belongs to an unrelated paper on structural-engineering models. The XR-1 paper is arXiv 2511.02776 — a single digit of the year, wrong in the model card, the tags and the suggested citation alike. Nothing on the pages has corrected it since publication.

▸Videos

No videos yet.

▸Reviews