HoloMotion (Horizon Robotics, v1.4, arXiv:2605.15336) is a motion-tracking foundation stack: a sparsely-activated MoE Transformer policy (60M→0.4B params, KV-cache inference at ~300 FPS on-robot), trained with distributed PPO on 80→2000+ hours of motion, with rough-terrain + medium domain-randomization defaults. Pipeline: raw datasets → HoloSMPL (canonical human H5) → HoloRetarget (robot refs) → HDF5 → PPO. Roadmap: v1 any-pose (done, G1), v2 any-command, v3 any-terrain, v4 any-embodiment (not built).
| GEAR-SONIC | HoloMotion | |
|---|---|---|
| Global root position | The historical report misread anchor_pos as XY drift; official eval checks anchor Z height | Reward weight 0.0; no drift termination at all |
| Root velocity | Absolute-error kernel, std 1.0 (a 0.15 m/s deficit is nearly free) | Relative-ratio kernel (error / (|v_ref|+0.1)) — a 20% deficit costs the same at any speed |
| Terminations | root height/orientation + ankle/hand height; primary removes the hidden full-XYZ foot term | timeout + fell-over (gravity projection 0.8) only |
| Headline metric | Success = never violate thresholds over full clip | MPJPE-global / MPJPE-local |
Separate horizontal speed probes remain useful for reference-data diagnosis, but the deleted evaluator does not show that SONIC officially terminated those clips or that the policy occupied a framework-induced local optimum. HoloMotion differs in reward design and remains worth studying, but those differences must be evaluated against the audited SONIC primary baseline rather than the withdrawn 38% narrative.
HoloRetarget is another Newton/Warp LM-IK retargeter, hard-coded to G1 (class HoloNewtonG1Retargeter; assets ship only unitree_g1). Porting it to X2 would face the same issues we documented with NVIDIA's soma-retargeter (straight-elbow singularity lives in the same newton.ik solver). However, it is bypassable: the robot-side H5 needs only ref_root_pos / ref_root_rot (xyzw) / ref_dof_pos — exactly what our existing GMR corpus (121k clips) and hybrid-graft pkls already contain. A pkl→H5v2 converter (~half a day) feeds our entire corpus in without touching HoloRetarget.
| Item | Work | Estimate |
|---|---|---|
| Robot config YAML | ~530-line spec: DOF, kp/kd, action scale, limits, asset paths — all values already known from the SONIC port | 1–2 days |
| Assets | URDF + MJCF exist; USD auto-converted by IsaacLab | 0 |
| Data | pkl→H5v2 converter, corpus reuse | 0.5 day |
| Training | PPO via accelerate, fits our 32-GPU setup; MoE-TF costlier per step than SONIC's MLP | runs |
| Deployment | Their real-robot stack is Unitree-only; X2 deployment remains ours (same as SONIC) | later |
Risks: young codebase (2026.04–07); embodiment transfer is a roadmap item, so X2 is uncharted there too; rough-terrain + DR defaults are a harder regime (may start flat); success-style guarantees weaker by design.