Archon Robotics releases Anchor 1 whole-body control model: report claims 28% lower wrist error; weights planned for staged release by end of 2026
According to a November 6, 2026 report by Machine Heart, Archon Robotics released the whole-body control model Anchor 1, positioned as a System 0 "cerebellum": in real-robot tests, global position errors at the wrists and ankles were about 28% and 39% lower than the previously best comparison method, and the success rate for collecting teleoperated kick-the-cardboard-box demonstrations rose from 37.5% to 98.0%; weights and deployment code are planned for staged release by the end of 2026.
What changed
According to a November 6, 2026 report by Machine Heart, the embodied AI company Archon Robotics released the whole-body control model Anchor 1, along with a public 32-page technical report and project page, positioning it as the System 0 (S0) layer in a whole-body intelligence architecture, i.e., the "cerebellum" that reliably turns commands into whole-body motion. The report says Anchor 1's commands specify only five body parts: the pelvis, left and right wrists, and left and right ankles, each with a target position and orientation, with other joints unspecified; deployment does not use online retargeting but instead performs a one-time calibration using the operator's and robot's T-poses, after which fixed scaling, rotation and offset parameters are applied frame by frame to convert human motion to robot scale. On the training side, retargeting is moved offline: each human motion segment is paired with a retargeted robot motion segment as supervision and reward, and after automated quality checks and manual review, the training set comprises 406,222 motion segments, about 1081 hours, of which 77% comes from self-collected motion-capture data. For correction, wrist and ankle tracking errors bypass the encoder and are fed directly to the decoder rather than through tokens, and a trajectory head is additionally trained so that target poses can be reconstructed when the upper layer provides only tokens. Real-robot tests were done on the Unitree G1: the report says that across 72 whole-body manipulation motions, global position errors at the wrists and ankles were about 28% and 39% lower than the best-performing comparison method; the success rate for collecting teleoperated kick-the-cardboard-box demonstrations rose from 37.5% to 98.0%, with successful demonstrations per hour rising from 9 to 43, and the slipper-picking task reached 62 per hour with a 97.37% success rate. The team also built three evaluation tracks, In-house, Public and Reachability, using a full-trajectory error threshold to judge whether a single motion segment succeeds. The coverage places this release in a broader context: in late January this year Figure added System 0 to Helix 02, in February Nvidia open-sourced the whole-body control foundation model SONIC, and since then Google DeepMind and AgiBot have also released related models. On the release plan, Archon Robotics says it will release Anchor 1's model weights and deployment code in stages by the end of 2026, along with teleoperation and demonstration data collection pipelines based on PICO and Noitom PN-LINK.
What it means for developers
For developers, the first thing to see clearly in this news is not a benchmark number but the shape of the interface. The two command paths described in the report — teleoperation or an upper-layer policy issuing five anchors, and an upper-layer policy directly outputting motion tokens — determine whether you are changing your data collection pipeline or your policy's output layer. If your existing stack does online retargeting at deployment, Anchor 1's approach is to move that step into offline training, with the deployment side passing only five anchors; the report says this needs only 60 milliseconds of future commands, whereas the compared approaches need 80 to 90 milliseconds, and it shows that four input devices — Xsens, CHINGMU, Noitom PN-LINK and PICO — can all be converted to five anchors, and operators from 165 to 188 centimeters tall and 50 to 80 kilograms can use it directly. Changes in collection efficiency carry through to cost: the reported rise from 9 to 43 successful demonstrations per hour means the human and robot time needed to reach the same volume of demonstrations could differ greatly, but this is Archon Robotics' own test result on the G1, not a general conclusion transferable to other bodies. The evaluation criteria also deserve attention: the coverage says that when the threshold is loosened, the success rates of the methods converge, and only when it is tightened to 10 centimeters does the gap appear, indicating that comparing controllers by whole-segment average tracking error can easily mask brief but fatal deviations. In addition, the coverage says the upper-layer policy can output only motion tokens, executed by a frozen Anchor 1, which changes the division of labor in "large model + controller." All of the above are claims from the report and the news coverage, with no third-party reproduction yet seen; before integration, the interface, latency and evaluation criteria should be verified as three independent variables rather than treating the report's ranking directly as a basis for selection.
Integration and evaluation
The following is conditional analysis and advice, not facts from the coverage. First, decide whether the input interface uses five anchors or keeps online retargeting: if the upstream is skeleton data such as Xsens, CHINGMU, PN-LINK or PICO and the downstream is a body like the G1, the path given in the report is to calibrate once, apply fixed parameters frame by frame, and no longer solve joint angles in real time at deployment; but if your robot's joint structure differs markedly from the operator's body shape, this calibration assumption needs to be re-verified yourself. A measurable criterion is: using your own reference motions, count the proportion of motions whose whole-trajectory wrist and ankle position error stays within 10 centimeters, compare with your existing online retargeting pipeline on the same batch of motions, and record the number of command jumps. Second, decide whether the interface exposed to the upper-layer policy is five anchor poses or motion tokens: if the upper layer is a large-scale VLA needing compact structured input, tokens are better for training; but if the scenario is teleoperated demonstration collection where precision is most sensitive, the report itself recommends issuing the five anchors directly, because reconstructing targets from tokens brings precision loss (in simulation, local position error rises from 21.9 millimeters to 22.4 millimeters). A measurable criterion is task success rate under a fixed number of trials, and the success-rate difference when the same task switches to the other interface. Third, decide on the evaluation criteria: switch from "whole-segment average tracking error" to frame-by-frame threshold judgments before comparing controllers, provided you already have a set of reference motions not used in training and accept process metrics such as slip counts and extra step counts; a measurable criterion is whether the success-rate gap between methods remains clear under the strictest threshold — if the gap disappears when the threshold is loosened, this evaluation lacks discriminative power for your current selection.
Limits and open questions
Both the coverage and the report list limitations. Evaluation was done on only one humanoid robot (the Unitree G1), with no strict cross-body comparison, so the above errors and success rates cannot be directly extrapolated to other models. The five anchors describe only where the body should go, not object geometry, contact forces or grasp stability; accurate motion tracking does not mean stable grasping; the success rate for picking up slippers during autonomous execution is 45%, and the coverage also acknowledges that autonomous whole-body manipulation is far from mature. The precision loss from reconstructing targets from tokens still needs to be reduced. The paper notes that methods compared on the Public Track may have encountered some public data during training, and the coverage never names that "mainstream open-source whole-body control model," so the reproducibility of the horizontal comparison is limited. On openness, the coverage mentions only a plan to release model weights and deployment code in stages, with no license terms stated, so "open weights" should not be treated as an open-source license. Weights and deployment code are planned for staged release by the end of 2026 and are not yet available; the capabilities, error and efficiency figures in this article are all vendor-reported. It should be clear that Anchor 1 does not currently appear in KimiSeek's model catalog, and the coverage mentions no hosted API; the release plans, availability timing and related pipelines mentioned in the article all refer to Archon Robotics' statements, not services already launched or provided by this site, and actual integration should follow the vendor's channels.