Archon's Anchor 1 controller lifts humanoid teleoperation from 9 to 43 good demos an hour on a Unitree G1

Archon Robotics (源策未来), a Shanghai embodied-AI startup founded in April 2026, released Anchor 1 on the evening of Nov. 5 Beijing time, a whole-body controller for humanoid robots that follows just five target points: the pelvis, both wrists and both ankles. In 72 whole-body motions run on a Unitree G1, the robot's global wrist position error was 75.69 mm, about 28% below the best baseline, MimicLite, at 105.74 mm, according to the project page. Ankle error fell about 39%, from 115.03 mm to 70.52 mm.

The number that matters more is in data collection. With an operator driving the G1 through teleoperation, the team collected 43 successful box-kick demonstrations an hour at a 98.03% success rate, against 9 an hour at 37.50% with the SONIC v1.1 controller. That is almost five times as many usable demos per hour. For picking up slippers, the rate rose from 41 to 62 an hour. Robot-learning teams pay for every hour of operator time, so a controller that throws away fewer takes cuts the cost of whole-body training data directly. That is the product Archon can sell first.

Five points instead of a full retargeted skeleton

The usual approach first retargets a human pose onto the robot's joints, then follows it. Anchor 1 skips online joint-space retargeting and takes the five anchor points as input, with a separate error-feedback signal telling the policy how far it has drifted, the paper says. It needs only 60 ms of future commands; the methods it was compared against need between 80 ms and nearly one second, Jiqizhixin reports. The training set holds 406,222 motions, about 1,081 hours, and 77% of them come from Archon's own motion capture.

Autonomy is where the gap is still open

When the same tasks ran without an operator, a policy using Anchor 1 kicked the box in 90% of 20 trials and picked up slippers in 45%, against 70% and 10% for SONIC v1.1, Jiqizhixin and the project page report. The slipper task more than quadrupled and still fails more than half the time. Anchor 1 also does not lead everywhere on the real robot: MimicLite's local wrist error, 25.63 mm, is lower than Anchor 1's 29.43 mm. The gains in that benchmark come in global position, where the robot's whole body has to end up in the right place.

Archon, which raised a seed round of several hundred million yuan in June from investors including ZhenFund, Gaorong Capital, IDG Capital and 5Y Capital, 36Kr reported, plans to release the model weights and deployment code in stages by the end of 2026, with PICO teleoperation and demonstration-collection tools, and to run a community whole-body control challenge in early 2027. The GitHub and Hugging Face links on the project page still read "available soon." Whether other labs can reproduce 43 demos an hour on their own G1s is the first test once they open.