Archon Robotics' Anchor1 cuts humanoid wrist tracking error to 75.69mm in real-world tests

机器之心 · wechat · 2026-11-06

Embodied-AI startup Archon Robotics released Anchor1, a whole-body control model for humanoids positioned as the "System 0" cerebellum layer, demoed by handwriting ANCHOR letters on a real robot and mimicking four motion-captured operators (165-188cm tall).

Design: commands specify only five anchors (pelvis, both wrists, both ankles); a one-time T-pose calibration yields per-limb scaling to retarget human motion without online re-direction (80.2% lower wrist trajectory deviation than parameterized-body-model inputs); tracking errors bypass the motion tokens and feed straight into the decoder as correction signals — ablations show removing this drops global success from 88.7% to 70.9%.

Data & eval: trained on 406k clips (1081 hours, 77% self-collected mocap) filtered by automated quality checks. The team also built a stricter three-track benchmark (In-house, Public, Reachability) with full-trajectory error thresholds. On a real Unitree G1, wrist/ankle global position errors were 75.69/70.52mm — about 28%/39% lower than the best prior method; teleop demonstration collection for a box-kicking task rose from 37.5% to 98.0% success, and from 9 to 43 successful demos per hour.

Context: after Figure's Helix S0, NVIDIA's SONIC, DeepMind and Zhiyuan, the field is shifting focus from "brains" to "cerebellums" and from demos to deployment. Paper: arxiv.org/abs/2611.00193.

Original post →

Archon Robotics' Anchor1 cuts humanoid wrist tracking error to 75.69mm in real-world tests

More from Embodied