Archon Robotics unveils Anchor 1 whole-body humanoid controller

Anchor 1 takes five target poses: the pelvis, both wrists and both ankles[1]. A geometric calibration maps the human motion to the robot's scale, and those targets are sent without online joint-space retargeting[1]. Archon names GMR and UMR as online retargeting methods whose poor solutions can hand the controller a robot reference that drifts from the intended human motion or changes abruptly. Those errors, Archon says, can then propagate into execution[1].

Training pairs each calibrated human sequence with an offline-retargeted robot motion of the same movement. The human poses are the commands. The robot motion guides the rest of the body[1]. The reference is encoded into tokens, and explicit tracking errors go separately to the control decoder, so the correction stays out of the motion encoder[1]. The controller builds on the SONIC motion-token formulation and adds a reconstruction head, so a higher-level policy can supply tokens directly[1].

The Paper uses a door for the same idea. A person's arm length and joint angles differ from the robot's, so the joint path cannot be copied. What matters is the hand's path to the handle, the centre-of-gravity shift, and the change in foot support[3]. Anchor 1, The Paper says, works with Archon's motion-control foundation model system, S0, to turn that intent into stable whole-body movement fitted to the robot[3].

The same calibrated interface supports optical motion capture, inertial motion capture and VR tracking, including PICO[1]. Archon shows four operators, 165 to 188 cm tall and weighing 50 to 80 kg, on upper-body movement, bending, squatting, kneeling and recovery[1].

Archon compares Anchor 1 with ScaleBFM, SONIC v1.1 and MimicLite on three tracks[1][2]. The in-house track has 15,342 motions. The public track has 2,230 motions from AMASS, BONES-SEED, CIRCLE and OMOMO. The reachability track has 3,210 references across standing, shallow squat, deep squat, staggered stance and half kneeling[1][2].

On the real robot, the project page and the paper both report 72 motions from the in-house track[1][2]. The paper also reports 51 standing reaches[2]. Archon says Anchor 1 shows lower global tracking errors and higher overall motion success across the position tolerances it tested[1].

The Paper says training used more than 400,000 real human motion clips, combined with simulation[3]. The paper's table is a different count: 406,222 sequences after left-right mirroring, approximately 1,081.4 hours, from in-house capture and public datasets[2].

For teleoperation, Archon compares SONIC v1.1 (SMPL) with Anchor 1[1][2]. Kicking a cardboard box, SONIC records 9 successful demonstrations an hour at 37.50% success, and Anchor 1 records 43 an hour at 98.03%. Picking up slippers, SONIC records 41 an hour at 68.33%, and Anchor 1 records 62 an hour at 97.37%[1][2]. Those Anchor 1 rows send errors to the decoder. The encoder-feedback variant records 45 an hour at 94.74% for the kick, and 67 an hour at 93.10% for the slippers[1][2].

Archon then trains vision-language-action policies on those demonstrations. They can command Anchor 1 with sparse body-part targets or with motion tokens[1]. Over 20 trials, box-kick success is 0.70, plus or minus 0.14, for SONIC with motion tokens, and 0.90, plus or minus 0.00, for Anchor 1 with reconstruction feedback and motion tokens. Slipper pickup is 0.10, plus or minus 0.00, against 0.45, plus or minus 0.07[1][2]. Commanded with motion tokens, the encoder-feedback variant scores zero on both tasks[2].

Model weights are expected in late 2026, and they are not published on the project page[1]. Deployment code, PICO teleoperation, demonstration collection and setup guides are expected in late 2026 as well. Leaderboards are in preparation, and an early access form is open[1]. Archon is preparing a community whole-body control challenge for early 2027, built around accurate target tracking, sustained coordination and stable support. It says the tracks, the submission details and the schedule will follow[1].

That 30% to 80% figure is Archon's result under the strictest real-world global scoring setting, with fewer look-ahead inputs and without retargeting[2]. The Paper's door example puts that interface in plain terms: the hand's path, the centre-of-gravity shift and the change in foot support, not a human joint angle the robot does not share[3]. The paper says the tests cover a single humanoid, and that tracking alone does not ensure a successful object interaction, because the sparse targets leave out object geometry, contact forces and grasp stability[2]. Weights and deployment code are still listed as upcoming[1].

Spectrum has covered a separate warehouse letter of intent, GXO and Apptronik's plan for up to 1,000 Apollo humanoids.