A robot reaches for an object on a shelf. Its wrist moves in roughly the right direction, but its feet shift and its body drifts. The motion still looks human. The grasp can already be lost. This gap between plausible movement and accurate interaction is the starting point of Archon’s Anchor1. [1] [2]
SONIC makes large-scale motion tracking a reusable control foundation. SuperMimic asks whether a learned differentiable dynamics model can make that foundation much cheaper to train. Anchor1 focuses on the execution interface: preserve the commanded wrist and foot geometry, then correct tracking errors without hiding them inside a motion token. These contributions sit at different points in the system.
Evidence checked through 14 November 2026: Anchor1 v1, SONIC v4, and SuperMimic v2. All experimental numbers below are author-reported. Anchor1 tests SONIC directly; it does not benchmark SuperMimic. Our architectural and industrial conclusions are interpretations, not an independent reproduction.
1. Compare the bottlenecks before comparing the scores
| System | Main intervention | Evidence to examine |
|---|---|---|
| SONIC | Scale motion data, policy capacity, and PPO training; unify command encoders through tokens | Held-out motion tracking, multiple interfaces, onboard runtime |
| SuperMimic | Learn shared dynamics; train the actor through short differentiable imagined rollouts | Compute-matched ablations and training-motion coverage |
| Anchor1 | Curate manipulation-oriented motion; separate sparse intent from explicit tracking feedback | Global versus local error, stationary reaching, demonstration yield |
These are whole-body execution systems. A controller tracking a commanded pose does not by itself identify an object, decide how to grasp it, or verify task completion. SONIC and Anchor1 can sit below a vision-language-action model (VLA); the high-level model and its training data remain separate dependencies. SuperMimic’s world model predicts low-dimensional robot dynamics during training, rather than generating a future RGB world. [2] [3] [5]
2. Anchor1’s important change is where feedback enters
Anchor1 first calibrates a human skeleton to robot dimensions and coordinate conventions. At runtime it consumes five target poses: pelvis, both wrists, and both ankles, including position and orientation. This leaves the controller freedom to choose the remaining joints instead of forcing a possibly poor full-body inverse-kinematics solution. However, offline-retargeted robot trajectories are still used in training and quality screening. Runtime retargeting-free does not mean retargeting-free data preparation. [2]
SONIC: shared control language
Anchor1: separate intent and error
SuperMimic: training loop
Conceptual synthesis from the papers; this figure does not describe a tested combined system
The encoder compresses the short reference window into finite scalar quantization (FSQ) tokens. Explicit wrist and ankle position/orientation errors bypass that encoder and enter the control decoder beside proprioception. A trajectory head reconstructs near-term targets; when a VLA predicts tokens directly, those reconstructed targets provide a feedback reference even though the motion encoder is bypassed. Longer-horizon prediction is an auxiliary training objective, not an extra future trajectory fed into the motor decoder. [2]
Our interpretation is that this separates a desired movement from the correction needed on this particular execution. If a latent command must also encode the current tracking error, its meaning changes as the robot moves. That is awkward for a high-level policy learning reusable commands. The paper’s token-interface ablation supports this concern, but reconstructed targets introduce their own error. The supplied-reference controller remains a stronger precision reference than assuming the two paths are interchangeable. [2]
The same controller has two reference-handling modes. Global mode needs root localization and preserves the initially aligned world trajectory. Re-locate mode repeatedly translates the reference window relative to the current robot root, avoiding global localization but allowing global drift. The hardware comparisons use re-locate mode without global-state input for any controller. Optical pelvis tracking and joint forward kinematics are used for scoring, not secretly supplied to the policy. [2]
3. Data scale hides very different collection costs
| System / corpus | Reported scale | Qualification |
|---|---|---|
| SONIC v4 | About 700 h acquisition; 611 h training split | Mirrored motions; unseen-content and unseen-repetition tests kept separate |
| Anchor1 | 406,222 sequences; 1,081.37 h after augmentation | 688.37 h original + 392.99 h mirrored; not 1,081 h of independent capture |
| SuperMimic main benchmark run | 264.56 h filtered BONES-SEED | Public motion references plus simulator transitions; not real-robot RL interaction hours |
Anchor1’s augmented inventory is 77.28% in-house data and 22.72% public data. The useful distinction is the behavior mix: reaching across postures, coordinating hands and feet, and keeping a tray level while moving. Automated checks reject floating, discontinuities, self-collision, jitter, and ground penetration; in-house motions additionally receive manual playback review. Mirrored sequences are screened separately. The result depends on capture, offline retargeting, and review labor, not just a sparse input vector. [2]
SONIC provides a useful counterweight: its evaluation explicitly separates unseen motion subcategories from new takes of known categories. SuperMimic’s headline failure count instead measures coverage of training motions. Anchor1 also constructs held-out public subsets, but its in-house benchmark should not be casually relabeled as wholly unseen behavior. A larger motion inventory, held-out action content, and useful task demonstrations answer different questions. [2] [3] [5]
4. What the direct Anchor1–SONIC comparison establishes
| Controller / tested look-ahead | Standing reach: wrist error | In-house motions: wrist error |
|---|---|---|
| Anchor1 / 60 ms | 44.58 mm | 75.69 mm |
| SONIC-G1 / 900 ms | 98.15 mm | 234.37 mm |
| SONIC-SMPL / 180 ms | 166.70 mm | 314.00 mm |
Source: Anchor1 Table 1, real Unitree G1, 51 standing-reach references and 72 in-house motion sequences. Values above are mean global wrist-position errors, not complete manipulation success rates. Global scoring retains displacement after the initial alignment, whereas local scoring removes each trajectory’s current pelvis XY translation. SONIC-G1 receives retargeted robot references; SONIC-SMPL receives human references. Interface and preview differences are part of this system comparison, not an isolated architecture experiment. [2]
Look-ahead is the amount of future motion required by the tested interface. It is not a forward-pass latency measurement. SONIC v4 separately reports 1–2 ms policy inference on Jetson Orin with TensorRT and CUDA Graphs, a 50 Hz policy loop, and 500 Hz motor-command streaming. Anchor1 reports a 50 Hz control rate and a 60 ms reference horizon; those figures alone do not establish lower measured inference latency. The SONIC repository also lists lower-latency SMPL checkpoints, so the tested 180/900 ms settings should not be generalized to every release. [2] [3] [4]
The cleaner architecture evidence comes from Anchor1’s own ablation. On its MuJoCo in-house benchmark, global-mode success rises from 70.9% without explicit feedback to 88.7% with decoder feedback and supplied targets. In re-locate mode, the latter configuration records 35.5% global success but 85.0% local success. This is a concrete illustration of why a locally plausible posture is insufficient when the task cares about world-relative placement. [2]
Here, success requires the entire motion to stay within wrist/ankle position and orientation limits of 10 cm and 1 rad, and a 25 cm limit for 14 body points. Mean errors use a separate early-termination rule and exclude the violating frame, so early failure can leave deceptively small averages. Simulation comparisons randomize physical parameters, but use deterministic sampling with seed 0, no external pushes, and no observation noise. They are not a comprehensive disturbance or safety test. [2]
5. SuperMimic changes the training bill, with a different claim
SuperMimic learns one dynamics model from simulator transitions shared across motions. During an actor update, model weights are fixed and gradients pass through short imagined trajectories; they never pass through the physical simulator. A critic estimates the return beyond the short horizon. Known simulator context, such as friction and calibration offsets, conditions the world model during training. The deployed actor uses noisy proprioception and a goal, rather than requiring those hidden parameters from a physical robot. [5]
| Coverage test: 129,514 training motions | Reported training budget | Failed motions |
|---|---|---|
| Released SONIC checkpoint | 128 L40 × 168 h ≈ 21k GPU-h | 179 |
| SuperMimic, early checkpoint | 1 H100 × 3 h = 3 GPU-h | 25 |
| SuperMimic, longer run | 2 H100 × 9.8 h = 19.6 GPU-h | 0 |
Source: SuperMimic v2, Figure 3 and Appendix F. Both methods have seen these benchmark motions during training. Every motion is replayed once from its first frame under deterministic physics, foot friction 1.0, no domain randomization, and no observation noise. Failure follows SuperMimic’s termination rule; each method retains its own actuator gains, action scale, and observation construction. The 0 is therefore training-set coverage in simulation, not zero failures on unseen motions or hardware. A separately tested SONIC v1.1 checkpoint fails 239 motions under that protocol. [5]
The compute difference is substantial, but GPU-hours are not interchangeable money or energy units: H100 and L40 differ, training corpora are 264.56 versus 611 hours, and the evaluations do not equalize the original training recipes. The stronger algorithmic evidence is the controlled LAFAN1 experiment. With matched motions, observations, rewards, termination, randomization, and actor/critic sizes, SuperMimic reaches normalized return 0.82 within 1.2 hours on one GPU. Simulator PPO does not reach it within 18 hours; even its eight-GPU run ends at 0.72. This supports a real efficiency gain without treating the cross-system headline ratio as a controlled cost reduction. [5]
Real G1 demonstrations include a skipping jump, a side roll, and flips, without per-motion tuning. They support physical transfer of the learned controller. They do not supply a matched Anchor1 comparison, precision-reaching statistics, or a long-running industrial failure rate. A cheap training method and a precise feedback interface could eventually be combined, but contact discontinuities, learned-model bias, and token reconstruction would need joint testing. That combination is our research hypothesis. [5]
6. The useful industrial metric may be accepted demonstrations per hour
| Collection setup | Kick box: accepted demos/h; SR | Pick up slippers: accepted demos/h; SR |
|---|---|---|
| SONIC v1.1 SMPL | 9; 37.50% | 41; 68.33% |
| Anchor1 decoder feedback | 43; 98.03% | 62; 97.37% |
These are real demonstration-collection results from Anchor1 Table 3, not autonomous policy success. A controller that avoids unwanted foot movement can reduce resets and make the operator’s data more useful. That is an economic argument with a measurable intermediate outcome. The evidence remains two selected tasks; the table does not establish a universal labor saving across operators, sites, or shifts. Capture hardware, review effort, retries, and downstream policy training must remain in the budget. [2]
Reproducibility currently favors starting with SONIC’s published code, checkpoints, and deployment documentation. Anchor1’s official page schedules weights, deployment code, and PICO teleoperation tools for late 2026; a schedule is not a release. For SuperMimic we verified the paper and browser demo, but the reviewed project links do not establish a complete downloadable training-and-deployment package. Anchor1’s public methods also do not provide a comparable total GPU-hour budget or end-to-end latency breakdown. These gaps prevent a complete three-way ownership-cost comparison. [1] [2] [4] [6]
Reference and data
Execution and scoring
Cost and acceptance
Proposed engineering workflow from 1N Robotics; not a protocol jointly validated by the three papers
7. Our judgment: follow the precision work, test the training work
Anchor1’s clearest contribution is bringing world-relative tracking and explicit execution feedback into a control interface that remains usable by humans and learned policies. It improves measured reaching and demonstration collection in the authors’ G1 tests. Its limits are equally concrete: one embodiment, reconstructed-target error, and sparse targets that do not specify object geometry, force, or grasp stability. A wrist can reach the right coordinate while the object still slips. [2]
SONIC remains a strong reusable baseline because scale, input diversity, held-out evaluation, and an available runtime come together. SuperMimic makes a credible case for changing the training engine, with its controlled PPO comparison more informative than its largest headline ratio. None of the evidence makes all three interchangeable, and none establishes unattended industrial operation.
For a humanoid team, our recommendation is to benchmark a repeatable reach-and-carry workflow now, separating reference quality from controller error. Test model-based training on a small matched task library before committing a large compute budget. Over the next one to two years, watch whether object-relative targets and force feedback improve complete manipulation, whether differentiable training retains its advantage under payload and contact shifts, and whether token-driven policies maintain precision during long tasks. Those results would matter more than another impressive motion montage.
Control precision closes only part of the learning loop
The UMI study starts at the other end: a useful reference must survive capture, reconstruction, and embodiment conversion before a tracker can execute it. GR00T 2 then illustrates policy-level task adaptation above a controller. These connections do not show that Anchor1 can replace SONIC inside GR00T 2 without retraining. The next comparison should separate reference error, tracking error, and object-contact failure on the same hardware.
Sources and version notes
- Anchor1 official research page (Archon / Yuance mirror)
- Anchor1 paper v1 — Tables 1–5, Sections 3–4
- SONIC paper v4 — evaluation, training, and deployment
- NVIDIA GR00T WholeBodyControl — releases and deployment
- SuperMimic paper v2 — controlled ablations and Appendix F
- SuperMimic official project and browser demo
The requested archon.tech research URL is also retained here as the original entry point; the accessible official Yuance page and paper supplied the verifiable details. SONIC v4 supersedes older compute figures; SuperMimic v2 explicitly states that its 129,514-motion headline is training-set coverage.