Compiled June 2026 · Focus: what makes a Vision-Language-Action model for a humanoid technically different from a tabletop-arm VLA — specifically (1) mobility / bipedal balance during manipulation, (2) the high-dimensional whole-body action space, and (3) data collection when you cannot simply teleoperate a balancing biped.

Companion reviews: System 0/1/2 architectures · GR00T series · Dexterous Manipulation · Cross-Embodiment · VLA Architectures · DuoCore-FS.

A tabletop VLA maps (image, instruction) → end-effector action. A humanoid cannot be modeled that way, for three reasons that organize this whole review:

1. The robot can fall over. Every action is conditioned on keeping a ~19–29-DoF underactuated biped balanced. Almost no system lets the VLA emit balance-critical torques directly — balance is delegated to a separate RL whole-body controller (WBC) and the VLA talks to it through a latent or a command.
2. The action space is 2–5× larger and heterogeneous. Arms + hands + waist + neck + legs, mixing position, force, and locomotion commands at different rates. The field's answers are latent action vocabularies, dual-system VLM→fast-policy splits, and stream-specific tokenization — not one monolithic head.
3. You cannot cheaply teleoperate a walking humanoid. So the data engine shifts off teleop toward human video → reconstruction → retargeting, motion-capture → RL-tracking → distillation, and sim-RL teacher → vision student. Teleop survives mostly for stationary upper-body work (Helix, GR00T's real layer).

The single most important structural fact: "humanoid VLA" today is a stack, cleanly described by the System 0/1/2 framing — a slow VLM (System 2, 1–10 Hz), a fast visuomotor policy (System 1, 50–200 Hz), and a reflex/balance layer (System 0, 500 Hz–1 kHz). The genuinely-new research is in how the VLA hands off to the balance layer and where the training data comes from.

Headline caveat — read before believing any "whole-body" claim: several flagship "humanoid VLAs" are upper-body only and never walk — Figure Helix (35-DoF, upper body, no paper, vendor numbers only) and GR00T's public GR-1 deployment (stationary bimanual). True bipedal loco-manipulation VLAs — where the policy commands walking/balance — remain rare and mostly simulation-trained: LeVERB, WholeBodyVLA, HumanVLA, Humanoid-VLA.

This is why humanoid VLA looks like "a VLM bolted onto a learned whole-body controller" rather than "scale imitation until it works."

The three-layer stack — a slow VLM hands off to a fast policy, which hands off to a balance layer:

flowchart TB
  VLM["System 2 · slow VLM · 1–10 Hz<br/>latent / subtask / motion code"]
  POL["System 1 · fast policy · 50–200 Hz<br/>diffusion / flow / AR head"]
  WBC["System 0 · reflex & balance · 500 Hz–1 kHz<br/>RL whole-body controller"]
  ROB["humanoid joints · 19–36 DoF"]
  VLM --> POL --> WBC --> ROB

The four families — which approach fits your constraint:

flowchart TB
  Q{Which family?}
  Q -- learned-latent VLA over an RL WBC --> F1[F1. Latent-vocabulary whole-body VLA<br/>LeVERB · WholeBodyVLA]
  Q -- VLM then fast-policy split --> F2[F2. Dual-system manipulation VLA<br/>GR00T · Helix · DuoCore-FS · Galaxea]
  Q -- language to motion / limited vision --> F3[F3. Language-to-motion controllers<br/>Lang-To-Loco · BFM-Zero · Humanoid-GPT]
  Q -- sim or human-video data engine --> F4[F4. Data & sim-to-real engines<br/>VideoMimic · VIRAL · EgoScale · Open-Sim-to-Real]

• F1 — Latent-vocabulary whole-body VLA. The VLA emits a learned latent action token; an RL WBC decodes it into balanced whole-body joint targets. The cleanest answer to all three axes at once. (LeVERB, WholeBodyVLA.)
• F2 — Dual-system manipulation VLA. Slow VLM + fast policy, heterogeneous DoF handled by per-embodiment MLPs. Strong on manipulation; balance is usually out of scope (upper-body or mobile-base). (GR00T N1–N1.7, Helix, DuoCore-FS, Galaxea G0, Fast-in-Slow.)
• F3 — Language-to-motion / promptable controllers. Bridge language (sometimes not vision) to whole-body motion via latents; these are often the System 0/1 substrate a full VLA sits on. (Lang-To-Loco/RoboGhost, BFM-Zero, Humanoid-GPT, HWC-Loco, HVD, LIFT, UniFP.)
• F4 — Data & sim-to-real engines. Not policies per se but the data-collection machinery that makes the others trainable. (VideoMimic, VIRAL, Open-Sim-to-Real, Gallant, EgoVLA, EgoScale.)

The defining humanoid problem. Three sub-patterns dominate:

The VLA almost never outputs balance torques. Instead a low-level RL WBC owns stability and the VLA steers it:

• LeVERB (arXiv 2506.13751, Berkeley): a System-1 RL WBC produces dynamics-feasible walking/turning/sitting from a latent instruction; the System-2 vision-language policy never touches joint torques.
• WholeBodyVLA (ICLR 2026, 2512.11047, AgiBot X2): a Loco-Manipulation-Oriented (LMO) RL policy for advancing / turning / squatting — "manipulation-aware locomotion" that moves the base to serve the manipulation goal. This is the explicit bipedal-loco-manip contribution.
• HumanVLA (NeurIPS 2024, 2406.19972): a state-based teacher trained with goal-conditioned RL + Adversarial Motion Prior (AMP) learns balanced walk-and-carry, then distilled into a VLA student.
• BFM-Zero (ICLR 2026, 2511.04131, Unitree G1, 29-DoF): a promptable Forward-Backward foundation controller with disturbance rejection (recovers from kicks/pushes never trained on) — the substrate a VLA prompts.

These are not VLAs but the balance layer the field depends on:

• HWC-Loco (ICLR 2026, H1/G1, 19-DoF): 3-stage hierarchy with a ZMP-feasibility constraint and robust optimization over an uncertainty set; survives ≤200 N pushes, 15 cm stairs, 20° slopes. Learned ZMP-threshold switching beats fixed thresholds (195 vs 459 switches).
• UniFP (CoRL 2025 Best Paper, 2505.20829): a single unified policy for position and force in legged loco-manipulation — infers external force from proprioception (no F/T sensor), no hand-engineered mode switching; +39.5% on contact-rich tasks (wiping, drawers).
• Gallant (CVPR 2026, 2511.14625): 3-D voxel-grid terrain perception (overhead + lateral constraints, not just a 2-D heightmap) for navigating pipes/shelves/stairs; near-100% staircase success.

Reasoning, control, and balance run ~100× apart: VLM 1–10 Hz → fast policy 50–200 Hz → balance 500 Hz–1 kHz. Industrial systems make this explicit — Figure's reported System-0 is a 1 kHz neural prior trained on >1,000 h of human motion capture that replaced ~109k lines of C++ balance code (per Review-System-0-1-2). The async-inference papers (DuoCore-FS 2.6×, Fast-in-Slow 117.7 Hz) exist precisely to keep the fast layer fed while the slow VLM lags.

Reality check: the most-publicized "humanoid" VLAs sidestep Axis 1 entirely. Helix is upper-body only (no walking/balance claim). DuoCore-FS explicitly omits a System-0 balance layer (quasi-static kiosk). GR00T N1's GR-1 deployment is stationary. Bipedal balance during VLA-driven manipulation is demonstrated by a short list — LeVERB, WholeBodyVLA, HumanVLA — and mostly in simulation.

DoF roughly doubles-to-quintuples vs a 7-DoF arm: H1 ~19, G1 ~23–29, Astribot S1 25, Figure 35 (upper only), Dexora 36 (bimanual+hands). Five distinct strategies:

Notes that matter:

• Explicit upper/lower-body decomposition is rare. Most systems use a shared latent and let projection layers sort out morphology. WholeBodyVLA's dual-stream head (separate arm-joint vs locomotion-command outputs) is one of the few explicit splits.
• Frequency mismatch is part of the action-space problem. Helix runs S1 at 200 Hz (80M params) under a 7–9 Hz 7B VLM; FiS reuses only the last 2 of 32 LLaMA blocks at high rate; DuoCore-FS runs an async 25–30 Hz fast head under a 1–3 Hz slow head.
• Dexterity rides on top. Bimanual 12-DoF hands (EgoVLA's Inspire hands, Dexora's XHAND) compound the DoF count — see Dexterous Manipulation for the hand-specific story; the whole-body papers mostly treat hands as another projected stream.

You cannot teleoperate a balancing biped at scale, so humanoid VLA data comes from four routes. Retargeting (human morphology → robot URDF under joint limits) is the recurring crux.

The dominant academic route — cheap monocular/egocentric human video lifted to robot actions:

• VideoMimic (CoRL 2025 Best Student Paper, 2505.03729, G1): everyday RGB video → 4-D reconstruction (SMPL mesh + scene, metric scale via joint Levenberg-Marquardt optimization) → retarget → 4-stage sim (mocap pretrain → scene-conditioned tracking → DAgger distill → under-conditioned RL). Removing the mocap-pretrain stage breaks it.
• EgoVLA (CVPR 2026, 2507.12440, H1 + 2×12-DoF Inspire): ~500K egocentric image-action pairs (HOI4D/HOT3D/HoloAssist/TACO), action = wrist SE(3) + MANO hand params, deployed via IK + hand retargeting. Pretraining lifts long-horizon success 2.22% → 45.93% vs ACT. Zero-shot (no robot finetune) = 0% — human video alone is insufficient.
• EgoScale (arXiv 2602.16710): 20,854 h of action-labeled egocentric video (monocular SLAM + 21-keypoint hand pose + CasADi/IPOPT retargeting) — ~20× prior ego-video efforts; feeds GR00T N1.7 with a log-linear scaling law L_val = 0.024 − 0.003·ln(D) (R²=0.998).
• WholeBodyVLA learns a Latent Action Model from action-free egocentric video — no demonstration labels — then a thin teleop layer grounds it.
• Humanoid-VLA (arXiv 2502.14795): language-motion pre-alignment → egocentric video-conditioned PEFT → self-supervised pseudo-annotation of unlabeled video.

For locomotion/whole-body motion (where balance, not semantics, is the target):

• Humanoid-GPT (CVPR 2026, G1): 2-billion-frame retargeted mocap (AMASS + LAFAN1 + Motion-X++ + MotionMillion + in-house), Harmonic-Motion-Embedding diversity clustering → per-cluster RL experts → DAgger into one causal GPT; zero-shot motion tracking at <1.5 ms (TensorRT).
• BFM-Zero: LAFAN1 (40 motions retargeted to G1), EMD-prioritized sampling, 192 M env-steps off-policy.
• Lang-To-Loco (RoboGhost): MotionMillion (50,378 seqs → 3,261 stable after filtering); language → 64-D motion latent → diffusion student; retargeting-free.

Privileged-state RL in sim, distilled to an RGB policy that transfers zero-shot:

• VIRAL (CVPR 2026, 2511.15200, NVIDIA/CMU/Berkeley, G1): privileged teacher → RGB student via large-scale tiled-rendering DAgger; compute is decisive (up to 64 GPUs; low-compute runs fail); 54 zero-shot real cycles.
• Open-Sim-to-Real (CVPR 2026, 2512.01061): staged-reset PPO teacher → RGB student → GRPO finetune for partial observability; humanoid door-opening 31.7% faster than human teleoperators.
• LeVERB and HumanVLA are entirely synthetic (rendered kinematic demos + RL); no teleop, no real video.

• Figure Helix: ~500 h multi-robot multi-operator teleop (vendor figure; no paper).
• GR00T: the "data pyramid" — 88 h GR-1 teleop (VIVE + Xsens) at N1, scaling to thousands of hours of real teleop by N1.6 (YAM/AGIBot/G1). The version-over-version story is which layer gets scaled: N1→N1.5 synthetic (DreamGen) · N1.5→N1.6 real-teleop · N1.6→N1.7 human-video (EgoScale).
• Dexora: hybrid exoskeleton backpack + Apple Vision Pro teleop (gross arm + markerless fingers); 100 K sim + 10 K real episodes.
• DuoCore-FS: just 10.22 h teleop, single kiosk task — efficient but narrow.

The data takeaway: academic whole-body VLAs lean synthetic / human-video; industrial manipulation VLAs lean teleop; GR00T is the only line that fuses all three at scale. Retargeting quality (metric-scale reconstruction, IK under joint limits) is the silent determinant of whether human-video data transfers.

• CoRL 2025 — human-video-as-cross-embodiment-source goes mainstream. VideoMimic (Best Student Paper) and UniFP (Best Paper) frame the year: monocular human video → balanced humanoid skills, and unified position-force loco-manipulation.
• NeurIPS 2025 — async dual-system inference. Fast-in-Slow embeds the fast policy inside the VLM's last blocks (117.7 Hz) — the architectural enabler for running a slow reasoner over a fast humanoid.
• ICLR 2026 — the whole-body controller substrate matures. A dense cluster of non-VLA controllers a VLA sits on: BFM-Zero (promptable FB foundation model), WholeBodyVLA (the standout true loco-manip VLA), Lang-To-Loco, HWC-Loco, HVD, LIFT.
• CVPR 2026 — the sim-to-real data engine. VIRAL, Open-Sim-to-Real, Gallant, Humanoid-GPT, EgoVLA — the vision venue owns how you generate humanoid training signal at scale (teacher-student, voxel terrain, mocap pretraining, ego-video).
• Industry (no papers) — dual-system + teleop. Figure Helix and NVIDIA GR00T define the deployed template; Helix is upper-body, GR00T is cross-embodiment with a real-teleop core.
• IROS 2026 🆕 — bimanual coordination structure + whole-body unification. The frontier is coordination architecture, not a second arm: 3D FlowMatch Actor (CMU/NVIDIA — one 3D policy for single and dual-arm, +41.4% PerAct2, beats 1000×-larger models) and EquiBim (bilateral symmetry-equivariance as an inductive bias). Whole-body: ULTRA (unified multimodal whole-body loco-manip), CEER (compliant EE+root unified interface), OmniDP (beyond-FOV omnidirectional 3D perception), DreamMimic (loco-manip via a world model), and a MoE-VLA for humanoid loco-manip (Responsibility-Induced Specialized Experts — the multi-task interference fix reaching whole-body). Data is the bottleneck → RL-based bimanual data generation + physics-informed retargeting (SPIDER). Context: IROS 2026 survey §5.3.

★ = Best/Best-Student Paper. "OR" = OpenReview ID (no arXiv listed).

1. Need true bipedal loco-manipulation (walk + manipulate)? → F1 latent-vocabulary VLA over an RL WBC (WholeBodyVLA, LeVERB). Accept that it's mostly sim-trained today.
2. Stationary/upper-body humanoid, want best manipulation? → F2 dual-system (GR00T, Helix-style). Balance is not your problem; invest in the fast head and teleop.
3. Balance/locomotion is the hard part, semantics secondary? → F3 controller substrate first (BFM-Zero, HWC-Loco, UniFP), then prompt it with a VLM.
4. Contact-rich loco-manip (push doors, wipe, carry)? → unified position+force (UniFP) under the WBC.
5. Data-constrained (no teleop farm)? → human-video / sim engine: VideoMimic or EgoScale-style ego-video for manipulation; VIRAL / Open-Sim-to-Real teacher-student for vision sim-to-real.
6. Can't hit real-time with a big VLM? → async fast-slow (DuoCore-FS, Fast-in-Slow) and a high-rate System-1 head.

1. VLA-driven balance is still delegated, not learned end-to-end. Every credible loco-manip VLA puts an RL WBC underneath; nobody robustly emits balance-critical whole-body torques directly from a language-conditioned policy. Whether end-to-end is even desirable is open.
2. The bipedal loco-manip VLA set is tiny and mostly simulated. LeVERB, WholeBodyVLA, HumanVLA — small sample, sim-heavy, few real-world long-horizon demos. Most "humanoid VLAs" are upper-body or mobile-base.
3. Data is the binding constraint, and retargeting is fragile. Human-video transfer hinges on metric-scale reconstruction + IK under joint limits; EgoVLA shows human video alone gives 0% zero-shot. No standard humanoid-VLA benchmark exists (LeVERB-Bench, Isaac Humanoid Manip Benchmark, RoboCasa are not unified).
4. Marketing ≠ capability. Figure Helix has no paper and is upper-body; vendor numbers are unverifiable. GR00T's later versions (N1.5+) are blog/model-card only, not peer-reviewed.
5. High-DoF credit assignment is unsolved. HVD's value-decomposition helps offline RL but the broad problem — learning coordinated 30-DoF whole-body actions from limited data — remains hard.
6. Compute walls. VIRAL shows vision sim-to-real for humanoids needs ~64 GPUs; this gates academic reproduction.
7. Frontier, unverified (cite with care): 2026 preprints surfaced but not source-verified here — PhysiFlow (2603.05410, multi-brain latent flow-matching whole-body VLA), HEX (2604.07993, humanoid-aligned experts for cross-embodiment whole-body), HumanoidExo (2510.03022, exoskeleton-data whole-body VLA), Cybo-Waiter (2603.10675). Verify before relying on them.

• ByteDance GR-2 / GR-3 (2410.06158 / 2507.15493) — strong generative VLAs but bimanual tabletop, not bipedal.
• Hume (2505.21432) — System-2 value-query VLA; general dexterous manipulation, not humanoid-specific.
• Dexora (2605.18722) — 36-DoF bimanual dexterous VLA; fixed-base, no locomotion (see Dexterous Manipulation).

• Companion reviews: System 0/1/2 · GR00T Series · DuoCore-FS · Fast-in-Slow · Dexterous Manipulation · Cross-Embodiment · VLA Architectures
• Whole-body VLAs: WholeBodyVLA · BFM-Zero · Lang-To-Loco
• Controllers / substrate: HWC-Loco · HVD · LIFT · UniFP
• Data / sim-to-real engines: VideoMimic · VIRAL · Open-Sim-to-Real · Gallant · EgoVLA

Verdict: the triple-system recipe (VLM + flow expert + RL lower body) became the open reference at RSS 2026; data efficiency, not data volume, is the winning argument — and evaluation lags arms by a generation.

RSS 2026 was humanoid loco-manipulation's coming-out: Ψ₀ (open foundation model; 800 h video + 30 h robot data beats >10× co-trained corpora incl. GR00T N1.6 by >40 pp), HiWET (world-frame EE tracking beats body-frame), HAIC (dynamics-aware interaction WM), EgoHumanoid + HoMMI (robot-free demonstration pipelines), plus a deep whole-body-control bench (TeleGate, OmniXtreme, X-Loco, pixel locomotion).

• Per-task fine-tuning sits inside every published loop (Ψ₀: 80 demos/task) — no humanoid zero-shot generalist.
• Payload and hand-DoF (Dex3-1-class) cap difficulty; precision insertion unsolved.
• No humanoid OOD protocol or perturbation suite; intervention-assisted scoring is the norm.

Latest (preprint): ω-0 🆕 — whole-body latent-predictive WAM for concurrent loco-manipulation (81.8% on 11 household tasks vs 44.5% ψ-0). See Latest Papers.

← Back to Home