Compiled April 2026 · Focus: how recent VLA papers (π-series + ICLR 2026 + CoRL 2025 + 2024–2026 arXiv) train one policy for many bodies, and how those designs compare.

This is a cross-paper review page organized by how the training data is mixed. For the deployment cut — which papers run one frozen checkpoint on multiple robots at inference (routed seen-robot vs zero-shot-to-unseen) — see the companion Single-Checkpoint Multi-Robot Deployment. For per-paper summaries see the linked Cross-embodiment pages. Related reviews: Review-VLA-Memory, Review-pi07, Review-VLM4VLA.

Every robot lab collects data on its own hardware. Every VLA needs to run on the hardware the customer has. The gap between those two is the cross-embodiment problem:

• Arm / gripper differences — different reach, DoF, force envelope, end-effector geometry.
• Base differences — fixed vs. wheeled vs. legged vs. humanoid, changes the whole workspace.
• Sensor differences — wrist cam vs. head cam; 3 views vs. 4.
• Control differences — joint space vs. end-effector space; 20 Hz vs. 50 Hz.
• Strategy differences — a bigger / heavier arm may need a completely different grasp strategy for the same task.

The headline 2026 result: π0.7 zero-shots laundry folding on a UR5e bimanual that was never in its laundry training data, matching expert teleoperators on their first attempt (85.6% task progress / 80% success vs. 90.9% / 80.6% for experts with 375 h avg experience). Strong evidence that embodiment transfer is real and emergent.

The counterpoint: the AnyBody benchmark (arXiv 2505.14986) shows that truly novel-morphology extrapolation (new link structures, new compositions) still fails for every method tested. In-distribution cross-embodiment works; out-of-distribution morphology does not — yet.

This review maps the design space between those two poles.

flowchart TB
  A[Cat A: Unified tokenized action space<br/>RT-X, Octo, RDT-1B, FAST]
  B[Cat B: Soft-prompt / body-conditioning<br/>X-VLA, HPT]
  C[Cat C: Embodiment-invariant latents<br/>UniVLA, UniAct, X-Sim, XR-1,<br/>UniSkill, TraceVLA]
  D[Cat D: Human video / wearable bridging<br/>DexUMI, EgoDex, Visual-Imit-Humanoid,<br/>ImMimic, UniSkill]
  E[Cat E: Morphology-aware architecture<br/>CrossFormer, HPT stems,<br/>GR00T N1, Octo, WholeBodyVLA]
  F[Cat F: Scale + prompt expansion<br/>π0.5 → π0.6 → π0.7,<br/>Gemini Robotics 1.5]
  G[Cat G: World-model-mediated<br/>DreamGen, Cosmos Policy, Ctrl-World]

  CHOICE{Your situation}
  CHOICE -- OXE-style data,<br/>simple baseline --> A
  CHOICE -- many robots,<br/>cheap new-robot extension --> B
  CHOICE -- human↔robot transfer --> C
  CHOICE -- data-starved --> D
  CHOICE -- legs / wheels / wings --> E
  CHOICE -- frontier lab data --> F
  CHOICE -- no real data for target --> G

§5 details each category with exemplars, pros, cons, and when to pick it.

• Mechanism: Unify 22 robots' data into one schema (~527 skills, ~160k tasks, 21 labs). RT-1-X and RT-2-X tokenize EE-pose deltas into discrete text tokens and train an LLM-style autoregressive policy on the union.
• Gap handled: Single-arm manipulators with different kinematics/cameras.
• Pros: Foundational dataset; positive transfer across 22 robots was a major 2023 result; the substrate every later VLA builds on.
• Cons: EE-pose-delta space is lowest-common-denominator; bimanual / dex / legged not well supported.
• Pipeline locus: Tokenizer (discrete EE-pose deltas).

• Mechanism: Transformer policy + diffusion action head; trained on 800k OXE trajectories to be a flexible initialization that fine-tunes to new sensory inputs and action spaces in hours.
• Pros: First truly open, downloadable cross-embodiment generalist; cheap fine-tune recipe.
• Cons: Per-robot specialists still beat it on hard tasks.
• Pipeline locus: Shared backbone + per-embodiment head swap.

• Mechanism: Single transformer with morphology-agnostic input handling — no hand-aligned obs/action formats. 900k trajectories across 20 embodiments including arms + wheeled base + quadcopter + quadruped.
• Pros: Matches specialists while spanning very different morphologies; first convincing "one net, many bodies" at the morphology level (not just many arms).
• Cons: Per-embodiment attention routing details modest; scaling behavior per-embodiment less characterized than X-VLA's.
• Pipeline locus: Unified tokenizer + shared trunk + per-embodiment readout heads.

• Mechanism: Stem / trunk / head factorization. Per-embodiment stems tokenize proprio + vision → fixed-size shared token sequence → shared trunk → per-task heads.
• Data: 52 datasets (real robots, sim, deployed, human video).
• Pros: Clean separation-of-concerns; +20% on unseen tasks; the precursor to X-VLA's "shared backbone + per-embodiment conditioning" idea.
• Cons: New robot = train a new stem (more machinery than X-VLA's single soft prompt).
• Pipeline locus: Embodiment-specific tokenizer (stem) + shared trunk + task head.

• Mechanism: Physically Interpretable Unified Action Space — explicit unified schema preserving physical meaning; 1.2B-param diffusion transformer specialized for bimanual manipulation.
• Pros: First serious 1B-scale diffusion foundation policy; strong bimanual; principled action-space unification.
• Cons: Unified action space lossy for dex hands; bimanual scope.
• Pipeline locus: Unified action space (training convention) + diffusion head.

• Mechanism: CoTracker extracts dense point trajectories from observation history; overlays active traces on the current frame; the VLA sees the overlay. Traces are embodiment-agnostic spatial-temporal context — they don't depend on arm appearance.
• Pros: +10% SimplerEnv, 3.5× real-robot vs. OpenVLA; no new parameters.
• Cons: 2D traces degrade under heavy occlusion; not an end-to-end universal action space.
• Pipeline locus: Input visual prompt (pixel-space).

• Mechanism: Vector-quantized vocabulary of universal atomic actions; a shared VLM outputs codes via Gumbel-Softmax; per-embodiment heads translate codes into robot-specific commands. 1M demos from 28 embodiments.
• Pros: A 0.5B UniAct reportedly outperforms 14× larger SOTA baselines on cross-embodiment evals. Explicit, interpretable atomic codes.
• Cons: Code count bounds expressiveness; fine-grained dex manipulation may not fit atomic codes cleanly.
• Pipeline locus: Intermediate representation (atomic action codes).

• Mechanism: Dual-system VLA — Eagle-2 VLM (System 2) + Diffusion Transformer (System 1). DiT blocks use an embodiment-aware state/action encoder that embeds robot-specific state + action.
• Data: Heterogeneous mix — real robot trajectories + human video + synthetic DreamGen rollouts.
• Pros: First open humanoid foundation model; tight sim/synthetic integration with NVIDIA's stack (Cosmos, Newton, DreamGen).
• Cons: Humanoid-leaning; embodiment-aware encoders still need per-robot calibration data.
• Pipeline locus: Embodiment-aware encoder inside DiT + VLM System 2.

• Mechanism: Hierarchical VLA: high-level subtask predictor + low-level flow-matching action expert. Co-training on multi-robot teleop + web VQA + object detection + semantic subtask prediction.
• Gap handled: Any arm the teleop mix covers (single-arm, bimanual, mobile).
• Pros: First open-world-home generalization at scale; template every subsequent generalist follows.
• Cons: Post-training often still needed for polish; no explicit cross-body mechanism — relies on scale + hierarchy.
• Pipeline locus: Scale + hierarchy.
• See: π0.5, π series evolution.

• Mechanism: Learn task-centric latent action tokens in DINO feature space via a latent-action model; humans + robots share this latent vocabulary; per-embodiment decoder maps latents → control.
• Pros: Beats OpenVLA with ~1/20 of pretrain compute and 1/10 of downstream data; works for navigation + manipulation.
• Cons: Bounded by DINO-feature quality; fine-grained contact may not fit the latent space; "task-centric" latents can collapse on visually busy, task-irrelevant video.
• Pipeline locus: Intermediate representation.
• Note: there are two ICLR 2026 papers called "UniVLA" — this is arXiv 2505.06111; the other is the Unified-VLA 8.5B at ICLR-2026-UniVLA.

• Mechanism: Extract object motion from human video → reproduce it in simulator → train image-conditioned diffusion policy → online sim-to-real adaptation. Sidesteps human-to-robot arm retargeting entirely.
• Pros: Principled; avoids the retargeting trap; clean recipe.
• Cons: Fails on tasks that aren't object-motion-reducible (force-sensitive insertion, gripper-geometry-critical); simulator fidelity is the ceiling.
• Pipeline locus: Intermediate representation (object motion) + sim-generated policy.
• See: X-Sim.

• Mechanism: Learn embodiment-agnostic skill embeddings from unlabeled cross-embodiment video (no aligned human-robot pairs needed); transfer to robot policies with few demos.
• Pros: Eliminates the expensive "aligned human+robot for same scene" data requirement.
• Cons: Encoder must separate skills visually; ambiguous actions with the same motion but different intent confuse the embedding.
• Pipeline locus: Intermediate representation (skill embedding).

• Mechanism: Wearable sensorized glove makes the human hand the universal manipulation interface. Record finger kinematics + contact → retarget to any robot hand.
• Pros: Breaks the dex-data bottleneck; captures natural grip diversity.
• Cons: Retargeting losses; some contact cues lost (palm, soft compliance); requires glove.
• Pipeline locus: Data-collection interface (front of pipeline).
• See: DexUMI.

• Mechanism: Not a method — a benchmark with 18 robot variations (8 procedurally generated + 10 real-world-based), two tasks (reach/push), with/without obstacles. Tests interpolation / extrapolation / composition over morphology.
• Findings: Zero-shot novel-morphology extrapolation and composition remain hard. In-distribution cross-embodiment works; genuinely new morphology does not.
• Why it matters: Canonical falsifier for "universal robot policy" claims. Every Group F method needs to confront this benchmark.

• Mechanism: 829 h of egocentric Vision Pro dex video, 194 tasks — a massive human-video dataset specifically designed to pretrain dexterous VLAs.
• Pros: Closes the data gap for Group D methods.
• Cons: Video only (no robot-action labels); retargeting is the user's problem.
• See: EgoDex.

• Mechanism: Train policies inside video-world-model rollouts; the world model abstracts embodiment dynamics at the pixel level.
• Pros: Infinite synthetic data; shared visual prior across robots.
• Cons: Pixel world models aren't yet precise enough for contact-heavy skills; heavy compute.
• Pipeline locus: World-model-mediated training data generator.
• See: DreamGen.

• Mechanism: Co-train on human video + small teleop set. DTW alignment maps retargeted human-hand poses to robot joints; MixUp interpolation creates intermediate human↔robot domains.
• Pros: Explicitly builds a bridge rather than hoping transfer emerges. Works across 4 distinct end-effectors (Robotiq, Fin Ray, Allegro, Ability).
• Cons: Quality of retargeting + DTW alignment is the ceiling; MixUp in action space is a strong assumption.
• Pipeline locus: Data-level.

• Mechanism: Internet video → simulated humanoid → real humanoid. Pose estimation + retargeting + tracking in sim + distillation into a contextual whole-body policy.
• Pros: First credible "humanoid skills from internet video" at scale; activity-appropriate postures emerge.
• Cons: Limited by retargeter, sim fidelity, and human ↔ humanoid mass-distribution differences.
• Pipeline locus: Data pipeline + sim training.
• See: Visual Imitation → Humanoid.

• Mechanism: Per-embodiment soft prompts (a handful of learned tokens per dataset) modulate a fully shared transformer backbone. New embodiment = train a new tiny prompt.
• Pros: Near-zero parameter cost per new robot; monotonic scaling law over # embodiments — the cleanest "scaling laws for cross-embodiment" story in 2026.
• Cons: Still needs demos per embodiment to fit the prompt; prompts are not interpretable; no explicit morphology topology.
• Pipeline locus: Prompt.
• See: X-VLA.

• Mechanism: Motion Transfer objective + novel architecture for learning from heterogeneous multi-embodiment data → zero-shot skill transfer across ALOHA ↔ bi-arm Franka ↔ Apollo humanoid without robot-specific post-training. Pairs with Gemini Robotics-ER 1.5/1.6 for embodied reasoning.
• Pros: Production-scale evidence of zero-shot cross-robot skill transfer; ALOHA → Apollo is a big morphology jump.
• Cons: Closed weights; mechanism details not fully public.
• Pipeline locus: Pretraining objective (Motion Transfer) + prompt/thinking stack.

• Mechanism: Dual-branch VQ-VAE (vision-dynamics branch + motion branch) with a shared codebook; downstream embodiments decode from the shared latent into their own action space.
• Pros: Strong inductive bias — forces an embodiment-free intermediate representation.
• Cons: Codebook quality is the bottleneck; fails if motion signals are too idiosyncratic (e.g., very different gripper kinematics); doesn't handle force/contact-only tasks.
• Pipeline locus: Tokenizer + shared codebook.
• See: XR-1.

• Mechanism: Fine-tune NVIDIA's Cosmos video foundation model with injected latent control tokens; model predicts next frame + action + value.
• Pros: Leverages industrial-scale video priors; interpretable future-frame predictions; supports value-based fine-tuning.
• Cons: Latents learned for generation, not control — precise low-level skills weaker; big inference cost.
• Pipeline locus: Backbone replacement (video foundation model) + token-level conditioning.
• See: Cosmos Policy.

• Mechanism: One VLA produces a unified latent that decodes to coordinated base / legs / arms / hands. Designed to capture loco ↔ manip correlation.
• Gap handled: Humanoid whole-body (not fixed-base).
• Pros: First credible "generalist-VLA recipe" for whole-body humanoid.
• Cons: Humanoid-specific; not designed for non-humanoid transfer; paired data is scarce.
• Pipeline locus: Unified low-level decoder off a shared latent.
• See: WholeBodyVLA.

• Mechanism: Prompt expansion rather than architectural cross-embodiment magic. Subtask text + subgoal images (from a 14B BAGEL world model) + episode metadata (speed, quality 1–5, mistake flag) + control-mode token (joint vs. EE) + per-component dropout at train + CFG at inference. MEM video-history encoder handles temporal context.
• Gap handled: Anything the data mix covers — single-arm, bimanual, mobile, partial humanoid, UR5e.
• Headline: Zero-shot UR5e bimanual shirt-folding = 85.6% progress / 80% success, matching expert teleoperators (90.9% / 80.6%) on their first attempt with that robot.
• Data: Multi-robot teleop + web + DROID + egocentric human video + RL rollouts + failures + autonomous rollouts (disambiguated by metadata).
• Pros: State-of-the-art production system; one model ≈ RL specialists; compositional generalization signal (unseen air fryer from 2 tangential episodes + web priors); discovers new grasp strategies on UR5e not present in training.
• Cons: Requires PI-scale data + metadata labeling; weights are proprietary; no head-to-head with ICLR 2026 discrete-diffusion VLAs.
• Pipeline locus: Prompt + data mix, zero architectural cross-embodiment magic. The UR5e result is arguably the strongest 2026 evidence that scale + prompting solves cross-embodiment for in-class morphologies.
• See: π0.7, Review-pi07, π series evolution.

ICRA 2026 (Vienna) adds a deployment-flavored layer to this map rather than a new paradigm; see the ICRA 2026 Survey. None of these claim novel-morphology extrapolation — consistent with AnyBody — and most stay firmly in-distribution.

• OmniVLA (nav) (UC Berkeley, 2509.19480) extends Category F into navigation: an ~8.26B OpenVLA-class backbone trained on ~9,500 h across 10 platforms, with randomized modality fusion over pose / goal-image / language goals. Its goal-modality dropout is the navigation analogue of π0.7's prompt-component dropout, and it reports zero-shot cross-embodiment transfer to Unitree Go1 and VizBot — a rare in-the-wild data point for Cat F outside manipulation. (Name collision: a separate manipulation "OmniVLA," 2511.01210, fuses IR/radar/audio and is unrelated.)

• Galaxea / G0 (2509.00576) is the cohort's counter-thesis to scale-driven Cat F: a 500 h / 100K-trajectory dataset on a single 23-DoF mobile-bimanual embodiment. Its three-stage curriculum cross-embodiment-pretrains (~1,000 h OXE) but finds the single-embodiment stage decisive, and in 20-trajectory few-shot tests embodiment-consistent pretraining beats no-pretraining. The takeaway sharpens Open Question Q1: raw cross-embodiment scale is not automatically better than consistent, well-annotated data for a fixed target.

• Toward Embodiment-Equivariant VLA Policy (2509.14630) attacks cross-embodiment generalization through equivariance rather than scale — a structural inductive bias that sits alongside Category E's morphology-aware architectures and is, in spirit, the architecture-not-data reply AnyBody invites.

• On the data/architecture side, OmniRetarget (2509.26633) is an interaction-preserving engine that augments a single demo across embodiments and terrains for humanoid loco-manipulation on a Unitree G1 — a Category D (retargeting) tool aimed at the humanoid whole-body regime — while DEFT (2602.02895) conditions a diffusion policy on embodiment and task so a damaged robot still completes it, a body-conditioning idea (Category B) repurposed for resilience rather than new-robot extension.

Net: ICRA 2026 mostly operationalizes existing categories (F omni-modal prompting, E equivariance, D retargeting, B body-conditioning) on real hardware, and Galaxea supplies the year's sharpest empirical caution against assuming more embodiments always helps.

• Mechanism: Discretize / bucket actions into one vocabulary → train LLM-style on everything.
• Exemplars: RT-X / OXE, OpenVLA, Octo, RDT-1B, FAST tokens / π0-FAST.
• Pros: ✅ Simple · ✅ leverages LLM training recipes · ✅ scales with data.
• Cons: ❌ Lossy for non-EE-pose actions (dex, humanoid) · ❌ lowest-common-denominator.
• Pick when: You have OXE-style data and want a strong baseline fast.

• Mechanism: Tiny per-embodiment tokens modulate a fully shared backbone.
• Exemplars: X-VLA (canonical), HPT (stems are a heavier cousin).
• Pros: ✅ Near-zero cost per new robot · ✅ clean scaling laws over #embodiments · ✅ backbone unchanged.
• Cons: ❌ Still needs demos per embodiment to fit the prompt · ❌ prompts not interpretable · ❌ no morphology topology modeling.
• Pick when: You have many robots and want the cheapest possible extension to a new one.

• Mechanism: Learn a latent that isn't about motors: object motion, skill embedding, trace, atomic action code, or a grasp manifold.
• Exemplars: X-Sim (object motion), UniSkill (skill embedding), UniVLA-2505 (latent actions), UniAct (atomic codes), TraceVLA (visual traces), XR-1 (UVMC shared codebook). NeurIPS 2025 addition: Grasp2Grasp (Princeton, 2506.02489) — treats cross-morphology grasp transfer as stochastic transport between grasp distributions via Schrödinger Bridges, with physics-informed costs on base pose / contact map / wrench / manipulability. Applies specifically to grasp-level transfer (not full trajectory) but generalizes across very different hands.
• Pros: ✅ Transfers across very different bodies · ✅ strong inductive bias · ✅ human↔robot transfer works.
• Cons: ❌ Latent quality is the ceiling · ❌ fine-grained contact often doesn't fit · ❌ ambiguous intents confuse latents.
• Pick when: You want clean human↔robot or multi-morphology transfer.

• Mechanism: Humans are the universal demonstrators; bridge via retargeting, wearables, or alignment.
• Exemplars: DexUMI (glove), EgoDex (Vision Pro), Visual Imitation → Humanoid (internet video), ImMimic (DTW + MixUp), UniSkill (video skill embedding).
• Pros: ✅ Unlocks massive unlabeled data · ✅ natural behavior diversity · ✅ no teleop bottleneck.
• Cons: ❌ Retargeting loss · ❌ force / contact often lost · ❌ sim fidelity caps the ceiling.
• Pick when: You are data-starved and the task is visible in human video.

• Mechanism: Explicit per-robot input/output heads or routing.
• Exemplars: CrossFormer (20 embodiments incl. wheeled / aerial / legged), HPT (stem/trunk/head), GR00T N1 (embodiment-aware DiT encoder), Octo (diffusion head with modular I/O), WholeBodyVLA.
• Pros: ✅ Respects per-robot physics · ✅ handles very different morphologies (legs, wheels, wings) · ✅ per-embodiment sensor configs are first-class.
• Cons: ❌ More parameters per new robot · ❌ harder to scale than Cat B.
• Pick when: Your "robots" include non-arm morphologies.

• Mechanism: Drown the gap in data; condition on rich prompts (subtask, subgoal image, metadata, control mode); let the model figure it out.
• Exemplars: π0.5 → π0.6 → π0.7, Gemini Robotics 1.5.
• Pros: ✅ State-of-the-art in practice — π0.7's UR5e zero-shot = expert teleop · ✅ minimal architectural baggage · ✅ absorbs Categories A, C, D as prompt modalities.
• Cons: ❌ Requires PI/Google-scale data & labeling discipline · ❌ opaque about what's doing the work · ❌ proprietary.
• Pick when: You have frontier-lab data access.

• Mechanism: Train policies inside a video world model that already abstracts embodiment.
• Exemplars: DreamGen, Cosmos Policy, Ctrl-World.
• Pros: ✅ Infinite synthetic data · ✅ shared visual prior across robots · ✅ new-robot deployment without real data.
• Cons: ❌ Pixel world models not yet precise enough for contact-heavy skills · ❌ heavy compute · ❌ sim-to-real still required.
• Pick when: You have no real data for the target embodiment.

1. 2023 — "can we even share?" RT-2 showed web VLMs help robots. Cross-embodiment was aspirational; teams hand-crafted per-robot stacks.
2. Late 2023 / 2024 — Group A (unified tokens) dominates. OXE + RT-1-X / RT-2-X demonstrate positive transfer across 22 robots. OpenVLA opens the recipe. Working assumption: "tokenize EE-pose deltas, pretrain on everything."
3. 2024 — Group E (morphology-aware architecture) appears. Octo, CrossFormer, HPT factor obs/action spaces by embodiment. The reply to RT-X: "we need architecture, not just data."
4. Early-mid 2025 — Group C (invariant latents) blooms. UniAct atomic codes, UniVLA latent actions, TraceVLA visual traces, X-Sim object motion, UniSkill skill embeddings. Thesis: "don't unify the action — unify something that doesn't depend on the body."
5. Mid 2025 — Group D (human video / wearable) mainstreams. DexUMI, EgoDex, Visual Imitation → Humanoid, ImMimic. π0.5 co-trains on web + human data. "The best cross-embodiment dataset is the human body."
6. Late 2025 — Group F (scale + prompt expansion) crowns production. π0.5 → π0.6 (Nov 2025) → π0.7 (Apr 2026). Gemini Robotics 1.5. Thesis: "give the model enough data + rich prompts and it figures out the embodiment by itself." π0.7 zero-shot UR5e is the headline.
7. Late 2025 / 2026 — ICLR 2026 synthesizes B + C + G. X-VLA = clean scaling laws via soft prompts; XR-1 = shared codebook across vision+motion; WholeBodyVLA = humanoid integration; DreamGen / Cosmos Policy / Ctrl-World = world models inside training.

Direction of travel:

• Mechanism consolidation. Soft-prompt (B) + invariant latent (C) + rich prompt (F) are converging — X-VLA's prompts and π0.7's metadata are the same idea from different angles.
• Data substrate unification. OXE + human video (EgoDex) + world-model rollouts (DreamGen / Cosmos) + RL traces (RECAP) is becoming the default pretraining mix.
• Benchmarks catch up. AnyBody (2025) makes novel-morphology generalization falsifiable — and shows we're not there yet on extrapolation/composition.
• Humanoid is in the tent. GR00T, WholeBodyVLA, Visual Imitation → Humanoid bring whole-body morphology into the same VLA discussion as tabletop.

Q1. Does scale alone solve cross-embodiment? π0.7's UR5e result suggests scale + prompting is enough for robots that "look like" your training distribution. But AnyBody shows extrapolation to new link structures still fails. ICRA 2026 adds a third caveat from the data side: Galaxea / G0 cross-embodiment-pretrains on ~1,000 h OXE but finds the single-embodiment stage decisive, and few-shot tests favor embodiment-consistent pretraining — i.e., raw cross-embodiment scale is not automatically better than consistent, well-annotated data for a fixed target. Updated answer: scale for coverage, architecture for extrapolation, and data consistency/quality for the deployment embodiment — three levers, not one. The "give it enough data and it figures out the body itself" thesis (trend #6) holds for in-class morphologies but is not a substitute for target-embodiment data quality.

Q2. Sim-to-real vs. real-to-sim for cross-embodiment?

• Sim-to-real (DreamGen, Cosmos Policy, Visual-Imit-Humanoid): best when task-completion is specifiable in sim and the embodiment gap is large (humanoid).
• Real-to-sim-to-real (X-Sim): best when human demos exist and tasks are object-motion-reducible.
• Real-only (π0.7, Gemini 1.5): best when frontier-lab data budgets are available.
• 2026 consensus: all three coexist; world-model-mediated sim (Group G) is closing the gap faster than expected.

Q3. Is there a universal action space that actually works?

• EE-pose-delta tokenized (RT-X, Octo) — works for single-arm, breaks on dex and humanoid.
• Physically Interpretable Unified Action Space (RDT-1B) — strictly better, still arm-centric.
• Atomic action codes (UniAct) — promising at coarse-skill level, unclear at contact-rich level.
• Latent action tokens (UniVLA-2505, XR-1 UVMC) — cleanest abstraction; tied to learned-feature quality.
• Empirical answer in 2026: no single representation wins across all skills. Production systems (π0.7) just condition on control-mode (joint vs EE) and let the flow-matching expert absorb the rest.

Q4. Can you transfer strategy, not just motion? π0.7 folding UR5e shirts reportedly discovers new grasp strategies not seen in training — strong strategy-transfer signal. UniSkill and UniAct also argue for strategy transfer via skill/atomic codes. But:

• Most cross-embodiment successes are within a class of strategies already in the data.
• Truly novel strategy invention (e.g., a parallel-jaw gripper discovering a pinch it has never seen) remains rare.
• The π0.7 UR5e result is the strongest 2026 evidence of genuine strategy transfer; academic reproducibility is the key question for 2026–2027.

Q5. Are benchmarks keeping up? AnyBody (2025), RoboArena ∞, RoboCasa365, WorldGym all push this direction. Expect a shakeout in 2026 where Group F claims get tested on AnyBody-style extrapolation suites — and we see which ones survive.

Q6. Closed vs. open ecosystems. The strongest cross-embodiment systems (π0.7, Gemini Robotics 1.5) are closed-weights. The strongest open alternatives (GR00T N1.6, OpenVLA, Octo, RDT-1B) trail on the hardest tasks. Whether open ecosystems close this gap in 2026 depends on who ships the next large multi-embodiment dataset — and whether it comes from academia (OXE v2?), industry (NVIDIA?), or a coalition.

If you're shipping a cross-embodiment VLA and wondering where to spend a quarter:

1. OXE-style data on a few single-arm robots? → Start with Octo / OpenVLA / RDT-1B as baselines (Cat A). Don't overthink it.
2. Many robots, want cheapest per-new-robot cost? → X-VLA-style soft prompts on a shared backbone (Cat B).
3. Big human-robot gap or human-video-abundant task? → Pick an invariant-latent method (Cat C — UniAct or X-Sim), or bridge via human demos (Cat D — DexUMI, EgoDex, ImMimic).
4. Legs / wheels / wings in the mix? → Go morphology-aware (Cat E — CrossFormer or GR00T N1).
5. Humanoid whole-body target? → WholeBodyVLA (Cat E) or GR00T N1.6 (Cat E+F) with paired loco-manipulation data.
6. Have frontier-scale data budget? → Follow the π0.7 / Gemini 1.5 playbook (Cat F): prompt expansion + massive multi-robot mix + RL rollouts distilled back in.
7. No real data for the target embodiment? → DreamGen or Cosmos Policy rollouts (Cat G) as a bootstrap; fine-tune on the real target once you have any data.

Don't skip benchmarking on AnyBody-style extrapolation. If your system only works on robots that look like the training distribution, that's fine — but the deployment story differs a lot from "universal."

• π series: π0.5 · π0.6 · π0.7 · π series evolution · Review-pi07
• ICLR 2026 cross-embodiment cluster: X-VLA · XR-1 · UniVLA · Cosmos Policy · WholeBodyVLA · Ctrl-World
• CoRL 2025 cross-embodiment cluster: X-Sim · DexUMI · Visual Imitation → Humanoid · DreamGen
• Data pages: EgoDex · RoboCasa365
• Companion reviews: Review-VLA-Memory · Review-pi07 · Review-VLM4VLA
• Surveys: ICLR 2026 survey · CoRL 2025 survey

Verdict: 2026 flipped the question — action-representation alignment is the precondition for cross-embodiment data to scale at all; hands joined the agenda with 80%+ zero-shot transfer.

Qwen-RobotManip's controlled scaling ablation is the pivot: naive zero-padded spaces yield no scaling law; a canonical 80-dim representation scales; camera-frame delta EEF scales log-linearly and transfers zero-shot to unseen arms (RoboTwin-XE 23.9% vs 14.5% joint; UR5 5.6× better). RSS 2026 extended to hands: DexGrasp-Zero (morphology-aligned graphs + fixed per-hand mapping, 85% zero-shot, +59.5% over SOTA) and One-Hand (canonical URDF + smooth morphology latent, 81.9%); LAP-style language-action pretraining and X-DiffVLA-class cross-embodied heads round out the cluster.

• Joint-space zero-shot transfer still fails (<5% on unseen morphologies).
• Morphology distance matters: Franka-class 7-DoF gaps resist even camera-frame transfer (5.9%); gripper↔dexterous-hand untried.
• Calibration is a hidden deployment tax for the best interface; uncalibrated fallbacks under-evaluated.
• Human↔robot is the hardest pair — see Review-Human-Video-Transfer.

← Back to Home