← Back to project

Humanoid Whole-Body Motion Tracking · Literature Review

curated for the UW humanoid team (X2/G1) · organized by lineage · 2026-07-23
literaturesurvey

1 · Physics-based imitation foundations (sim)

WorkVenue/YearOne-line takeawayRelevance to our X2 effort
DeepMimic (Peng et al.)SIGGRAPH 2018RL imitation of mocap clips with reference-state init & early termination — the founding recipe.RSI/ET machinery we use daily descends from this.
AMP (Peng et al.)SIGGRAPH 2021Adversarial motion priors replace hand-designed imitation rewards.Alternative reward paradigm if explicit tracking rewards keep fighting us.
PHC / PULSE (Luo et al.)ICCV 2023 / ICLR 2024Perpetual simulated humanoid tracking of ~all AMASS; distilled latent skill space.The sim-side ceiling proof: near-universal tracking is possible with enough data+compute.
MaskedMimic (Tessler et al.)SIGGRAPH Asia 2024Masked motion inpainting unifies tracking, VR, text control in one physics policy.Blueprint for multi-mode conditioning (SONIC's 3-encoder design echoes this).

2 · Real-robot whole-body tracking & teleop

WorkVenue/YearOne-line takeawayRelevance to our X2 effort
H2O / OmniH2O (He et al.)IROS/CoRL 2024Real-time human-to-humanoid teleop via RL tracking + retargeted refs.First proof the sim recipe survives real G1-class hardware.
HumanPlus (Fu et al.)CoRL 2024Whole-body shadowing + autonomous skills from human data.Data-flywheel story for humanoid skills.
ExBody / ExBody2 (Cheng et al.)RSS 2024 / 2025Expressive whole-body control; decoupled upper/lower tracking targets.Upper/lower decoupling mirrors our SOMA-lower + GMR-upper graft.
HOVER (NVIDIA)ICRA 2025Distills multiple control modes into one neural whole-body controller.Multi-mode distillation = likely future step after tracking works.
ASAP (He et al.)RSS 2025Aligns sim and real physics via residual action models for agile skills.The sim2real delta-learning we may need for X2 deployment.
TWIST (Ze et al.)2025Teleoperated whole-body imitation system on real humanoids.Same GMR author lineage; teleop-grade retargeting constraints.
BeyondMimic / KungfuBot2025Dynamic/agile motion tracking pushed to highly dynamic skills on real robots.Upper bound on what tracking-based control achieves on hardware.

3 · Scale-up: foundation-style controllers

WorkVenue/YearOne-line takeawayRelevance to our X2 effort
GEAR-SONIC (NVIDIA)arXiv:2511.07820, 2025Universal motion tracking at scale (8k+ hrs, multi-GPU): ~99% success on G1; 3 conditioning encoders.A historical released-default common-v4 replay scored 104/126 = 82.54%; the audited primary 50k result is 480/512 = 93.75%, and the old 38% figure was a mislabeled custom XY/3D diagnostic.
HoloMotion-1 (Horizon Robotics)arXiv:2605.15336, 2026MoE-Transformer motion foundation model; zero-shot tracking; video-reconstructed + mocap + in-house data; HoloRetarget 3000+ FPS; HoloSMPL unified capture.Candidate independent comparison framework — evaluation in progress.
GR00T N1/N1.5 (NVIDIA)2025VLA stack whose action layer is a decoupled whole-body controller.Where tracking controllers plug into task-level autonomy.

4 · Retargeting & data production

WorkVenue/YearOne-line takeawayRelevance to our X2 effort
GMR (Ze et al.)2025Real-time general motion retargeting, per-frame differential IK, any URDF.Our production retargeter; no contact awareness by design (teleop origins) — root cause of our feet defects.
SOMA Retargeter (NVIDIA)2026Offline trajectory LM-IK + feet stabilization + limit clamping on Newton/Warp; G1-only public configs.Dissected in our postmortem: lower body excellent on X2, upper body numerically unstable (elbow singularity, a “1-DOF waist” attribution later found incorrect — X2 has a 3-DOF waist).
OmniRetarget2025Interaction-preserving retargeting for loco-manipulation data.Relevant when we add object interaction clips.
SEED dataset (Bones Studio)HF 2026Large SOMA-skeleton mocap library + official G1 retargets (SOMA-produced).Our source corpus (122k clips); SOMA-BVH archive downloaded (45.5 GB).
AMASS / SMPL-XICCV 2019Unified mocap archive / expressive parametric body model.Upstream of every retargeting pipeline here.

5 · Our internal reports

Team doc (working notes): UW Humanoid Research — Google Doc

literature review · maintained alongside the X2 project · 2026-07-23