0. Humanoid Commercialization Cadence (Jun 2026) — Active

Status: From public spectacle to paid deployment and hard shipment volume | Key sources: TrendForce shipments, Figure 03 / BotQ / BMW, Optimus Boston Marathon, Beijing Half-Marathon, Figure 03 / BMW scaling Key players: Tesla, Boston Dynamics, Figure AI, 1X, Unitree, AgiBot, Booster Robotics, Fourier

The April public-visibility wave (Beijing E-Town 21.1 km half-marathon Apr 19; Optimus's first US uncontrolled appearance at the Boston Marathon Apr 21) has by mid-2026 hardened into commercial and volume signals:

• Hard shipment volume (TrendForce, Apr 9): global humanoid shipments forecast to breach 50,000 units in 2026 (+700% YoY). China output grows up to 94%; Unitree + AgiBot capture ~80% of Chinese shipments. Unitree commits to 75K humanoid + 115K quadruped/yr capacity; AgiBot scaled 1,000 → 10,000 units (Expedition A3, Mar 2026). The global supply base is overwhelmingly Chinese.
• Paid commercial deployment (Figure, Apr 29): Figure 03 in paid deployment at BMW Spartanburg (~40 units) across body-shop/assembly; BotQ output 1/day → 1/hour (24x in <120 days), 350+ delivered. The horizontal-platform thesis now has revenue, not just a pilot.
• Public-market test (Unitree, Jun 1): Shanghai Stock Exchange listing committee cleared Unitree's IPO — first humanoid pure-play approved for China's A-share market (~$6.2B valuation, ~$616M raise).
• Tesla V3 timeline (mid-2026): Optimus V3 production confirmed for summer 2026 (high volume summer 2027); Model S/X Fremont lines retired Q2 2026 to convert for Optimus.

The three go-to-market models are now empirically distinguishable: horizontal platform (Figure — paid OEM labor), vertical self-deployment (Tesla, Hyundai/Boston Dynamics), and consumer subscription (1X NEO). All three are simultaneously live, but only Figure has demonstrated third-party paid revenue at fleet cadence.

What to watch: Tesla Optimus V3 summer-2026 volume vs guidance. Whether Figure adds OEM partners beyond BMW. Unitree IPO pricing/aftermarket and whether other Chinese makers (AgiBot) follow to public markets. Whether the ~50K shipment forecast is met and how much is productive deployment vs research/quadruped units. Western programs closing the volume gap with China (capital lead ≠ shipment lead).

---

Research Frontier: Robotics & Humanoid Automation

What's genuinely new and where the field is heading.

Active Frontiers

1. Zero-Shot Loco-Manipulation via Foundation Models

Status: Rapid progress Key papers: Humanoid-COA Key players: Unitree, NYU, Harvard, UCL

Humanoid-COA demonstrates that vision-language models (GPT-4V) can decompose natural language instructions into executable whole-body behaviors without task-specific training. 96.6% grasping, 90% mobile pick on physical robots. This is the "ChatGPT moment" for humanoid control — foundation models as the reasoning layer, pre-trained controllers as the execution layer.

The ACM survey (Cao 2024) frames this as the transition from the "human-looking" to "human-like" paradigm — behavioral correspondence with human intent enabled by GenAI, not just physical resemblance. VLA (vision-language-action) modeling is the emerging next step: unified models that jointly process visual scenes, instructions, and action histories to generate real-time motor commands.

Open problems:

• Long-horizon combined tasks still 56-63% success
• Dependence on external APIs (latency, availability)
• Recovery from mid-task failures
• VLA training data requirements and generalization at scale

2. Sim-to-Real at Production Scale — and into Dexterity

Status: Rapid progress; dexterous gap now closing Key papers: ABB + NVIDIA HyperReality, DexSim2Real, Sim-to-Real Gap of Foundation Model Agents Key players: ABB Robotics, NVIDIA, Physical Intelligence / dexterous-RL labs

99% sim-to-real correlation was the positioning milestone — robots trained entirely in simulation deploy to production lines with minimal debugging, via ABB's identical virtual/physical firmware plus NVIDIA's deliberate sensor-imperfection injection. The 2026 advance is the dexterous milestone: DexSim2Real (Zeng et al., May 2026) attacks contact-rich manipulation with a VLM-as-realism-critic for domain randomization (FM-DR), tactile-visual cross-attention (TVCAP), and an LLM-decomposed curriculum (PSC), reaching 78.2% real-world success across six tasks with the sim-to-real gap cut to 8.3% — beating DrEureka and DeXtreme. The gap that ABB explicitly left open is now narrowing, though dexterity remains an order of magnitude looser than positioning.

The concept is also generalizing beyond robots: a KDD 2026 position paper (Liu et al.) argues LLM-agent robustness is itself a sim-to-real problem over the four MDP elements (Observation/Action/Transition/Reward) and should adopt robotics' domain-randomization toolkit — inverting the usual ML→robotics borrowing.

Open problems:

• Does 99% correlation hold for dexterous manipulation (not just positioning)? — Partly answered: DexSim2Real reaches 8.3% dexterous gap, but only on six tasks and far from positioning-grade.
• Deformable object handling in simulation
• Sim-to-real for contact-rich tasks (assembly, cooking)
• Do the four MDP-gap categories yield standardized stress-test benchmarks transferable between robot policies and LLM agents?

2b. Humanoid Foundation Models — Recipe Fragmentation

Status: Rapid progress, recipe battle intensifying Key papers: Ψ₀, π₀.₅, Humanoid World Models, GR00T N1.7, Xiaomi-Robotics-1, ACE-Brain-0.5, S²-VLA Key players: Physical Intelligence (π₀ lineage), NVIDIA (GR00T), Meta FAIR (V-JEPA 2-AC — see ai KB), Xiaomi

Three distinct recipes competed for the "humanoid foundation model" crown as of mid-2026, and all four key entrants published or shipped within a 12-month window. Physical Intelligence bets on heterogeneous co-training (π₀ → π₀.₅). NVIDIA GR00T bets on large-scale egocentric human video (20K hours) + ecosystem integration (Isaac Sim). Ψ₀ bets on extreme data efficiency (800h human + 30h robot beats 10× more data). V-JEPA 2-AC bets on passive video pre-training + minimal action adapter.

A single week in July 2026 (Jul 1-16) added three more distinct bets rather than converging the race: Xiaomi-Robotics-1 pushes data-scale an order of magnitude further than GR00T (100k+ hours of UMI-collected trajectories, auto-labeled) and reports the scaling is still unsaturated — new SOTA on RoboCasa365 (57.4% vs. 46.6%) and RoboDojo (20.07 vs. 13.07); it is also the first major consumer-electronics company (not a robotics lab or hardware maker) to enter the race directly. ACE-Brain-0.5 bets on modular function-unification — five coupled functions (perception/decision/interaction/monitoring/self-improvement) behind one 8B backbone via a new SSR+ recipe, improving 14/18 spatial-perception benchmarks but explicitly "competitive" (not best) on navigation/manipulation. S²-VLA bets on architecture over scale — a 2B-parameter model with adaptive-fusion attention beats 7B-scale baselines on long-horizon manipulation, directly answering this KB's own "long-horizon combined tasks still 56-63% success" open problem (area 1) with an architectural fix rather than more data. Full detail in Foundation Models for Robotics.

Open problems:

• Which recipe wins — co-training (π₀.₅), data-scale (GR00T, Xiaomi-Robotics-1), data-efficiency (Ψ₀), passive-video (V-JEPA 2-AC), modular unification (ACE-Brain-0.5), or adaptive-fusion architecture (S²-VLA) — or do they specialize by deployment regime?
• Can humanoid foundation models hit product-market fit before Tesla Optimus / Figure scale vertical integration renders the "open model" path commercially moot?
• Cross-embodiment transfer — does a model trained on one humanoid transfer to another?
• Does S²-VLA's architecture-over-scale result generalize beyond LIBERO/SimplerEnv simulation benchmarks to real-robot deployment?
• Now that Xiaomi has entered on data-scale alone, does the recipe race concentrate around 2-3 dominant approaches, or does it keep fragmenting as more well-resourced entrants join?

2c. Commercial Humanoid Deployment — 2026 Inflection

Status: Inflection year for commercial scale | Key sources: TrendForce shipments, Figure 03 / BotQ / BMW, Tesla Optimus Gen 3, Boston Dynamics Atlas Now grounded (mid-2026):

• Volume floor exists — TrendForce: 50,000 global humanoid units in 2026 (+700% YoY); China +94%; Unitree + AgiBot ~80% of Chinese shipments. The survey-projected $38-243B-by-2035 market has a concrete 2026 floor.
• Paid third-party deployment exists — Figure 03 at BMW Spartanburg (~40 units, paid), BotQ at 1 robot/hour. The horizontal-platform model has revenue.
• Public-market validation exists — Unitree's Shanghai IPO cleared (Jun 1); first humanoid pure-play on a major exchange.
• Boston Dynamics Atlas — production launched at CES 2026 (Jan 5); all 2026 units committed to Hyundai RMAC and Google DeepMind; 30K-unit/year factory planned for 2028.
• Tesla Optimus — 1,000+ deployed; V3 production summer 2026; Fremont Model S/X lines retired for Optimus; Giga Texas 10M/yr facility under construction.
• 1X NEO — Q2 2026 consumer delivery; $20K or $499/month.

Why this matters: Commercial-scale deployment arrived in 2026 faster than most forecasts projected, and mid-2026 added the three things April lacked: shipment volume, paid revenue, and a public listing. Talent, capex, and policy attention follow these commitments — and the supply base is overwhelmingly Chinese.

Open problems:

• Will OEM humanoid makers license foundation models (GR00T, π₀.₅) or build in-house?
• When does the $20K NEO / Optimus price target actually hold at unit economics?
• How much of the ~50K shipment figure is productive deployment vs research/demo/quadruped units?
• Which industrial vertical has first real ROI — automotive (Hyundai/BMW), logistics (Amazon), or manufacturing (Tesla)?

3. Consumer Humanoid Robots

Status: Early stage, high momentum Key papers: 1X NEO World Model, Figure 03 + Helix 02 Key players: 1X Technologies, Figure AI

Two companies are converging on consumer humanoids in 2026: 1X (NEO at $20K, Q2 delivery) and Figure AI (Helix 02 for household tasks). Both use teleoperation/mocap data to bootstrap, then scale via simulation and progressive autonomy. The White House demo signals political legitimacy.

The ACM survey projects the humanoid market at $38–243B by 2035 (13.8–50% CAGR). The wide range reflects uncertainty about whether consumer segment capabilities — requiring the "human-like" paradigm — will be achieved this decade.

Open problems:

• Safety in unstructured home environments
• Economics of consumer pricing ($20K is aspirational, $499/mo may be more realistic)
• Task generalization beyond demonstrated capabilities

3b. Embodied-AI Inference Serving — the Compute-Demand Layer (New, Jul 2026)

Status: Early stage — single paper, but structurally important Key papers: ROSA Key players: NVIDIA (co-authors), Stanford, Xiaomi/GR00T/π₀.₅ as the models being served

This is a genuinely new frontier area for the KB, opened by this sweep and given its own concept page: Embodied-AI Inference Serving & Compute Economics. ROSA (Jiang et al., Stanford/NVIDIA, Jul 2026) is the first systems paper to treat robot-fleet inference serving — not training — as a first-class compute-economics problem, importing the "disaggregate compute from the physical unit" logic that already transformed LLM serving (vLLM, disaggregated prefill/decode). Instead of one edge GPU per robot, ROSA pools server-class GPUs across a fleet and schedules against a factory-productivity objective (not per-request latency), reporting up to 12.06x factory-productivity improvement over conventional dedicated serving.

This matters structurally, not just as one paper's benchmark: it is the first concrete evidence that robotics foundation models are now large/deployed enough that inference-serving infrastructure is a distinct, separate compute-demand curve from training-time scaling (area 2b above). As fleets scale toward TrendForce's ~50,000-unit 2026 floor (area 2c) and models keep growing (Xiaomi-Robotics-1's unsaturated 100k+ hour scaling, ACE-Brain-0.5's 8B backbone), the two curves — train-time and serve-time — compound rather than substitute for each other.

Open problems:

• ROSA is a single preprint, not yet independently corroborated — does the 12.06x figure hold outside its own benchmarked workloads and embodiments?
• What happens to shared-pool serving gains under real factory network failure conditions (unaddressed in the source)?
• Does shared-pool serving generalize to non-factory (consumer/home) humanoid deployments that cannot assume a local server-class GPU pool?
• Is there a fleet-size threshold below which dedicated per-robot edge GPUs remain cheaper than shared-pool serving?

4. World Models for Robot Learning

Status: Active frontier — industrial and research tracks converging Key papers: 1X NEO World Model, V-JEPA 2, H-WM, StructVLA, Wayve GAIA-2 Key players: 1X Technologies, NVIDIA, Meta FAIR, Google DeepMind, Wayve Cross-topic: See ai/wiki/concepts/world-models.md for the full research picture.

Two parallel tracks are converging. Industrial: 1X's NEO world model enables environmental understanding and self-directed skill acquisition; NVIDIA's Isaac Sim 5.1 creates high-fidelity simulated worlds with deliberate sensor imperfections. Research: Meta FAIR's V-JEPA 2 achieves zero-shot Franka pick-and-place after passive video pre-training plus <62h of robot video; H-WM enables long-horizon TAMP via hierarchical symbolic+visual prediction; StructVLA rejects dense pixel rollouts for sparse structured keyframes. Wayve's GAIA-2 is the commercial AV parallel — generative world models in production for sim-to-real training.

The field has split into two architectural camps — JEPA (abstract-representation prediction, favors control) and generative (pixel-space prediction, favors simulation/data augmentation) — with physics-consistency benchmarks (PhyWorldBench, VideoScience-Bench) showing generative models at 58-64% on phenomenon congruency, catastrophic for control but tolerable for simulation. The convergence of world models + sim-to-real could eliminate the need for per-task human demonstrations.

A late-June 2026 result complicates the clean split: Wang et al. adapt Cosmos Policy — a video-diffusion (generative-camp) world model — to act directly as the controller itself, not just as a simulator, trained on ~800 purely-synthetic demonstrations per task with zero real-world data, and deploy it zero-shot on a physical Franka arm at 35% average success (the first reported sim-to-real transfer of a world-action model; see Sim-to-Real Transfer). 35% is well below JEPA/architecture-native results elsewhere in this KB (DexSim2Real 78.2%, S²-VLA SOTA), so it is consistent with — rather than a refutation of — the "generative = simulation-grade" specialization, but it is the first data point testing that specialization against a generative model actually deployed as a controller rather than only as a simulator.

Open problems:

• Does V-JEPA 2-AC's tabletop success transfer to long-horizon, multi-step manipulation?
• Scaling world models to unstructured home environments (where 1X is betting)
• Real-time inference constraints on robot hardware
• Grounding predictions in physical dynamics (not just visual patterns) — the physics-consistency benchmark gap
• Can the industrial track (1X NEO, Isaac Sim) and the research track (V-JEPA 2, H-WM) merge, or stay parallel?
• Does the teleoperation-to-autonomy pipeline (1X, Figure, Tesla) out-scale world-model-based approaches, or do they become complements?
• Does Cosmos Policy's 35% real-world success as a generative-model-as-controller close toward architecture-native/co-trained-VLA reliability with more synthetic scale, or does zero real-world feedback impose a structural ceiling?

5. Tactile Sensing for Dexterous Manipulation

Status: Rapid progress Key papers: Tactile In-Hand Rolling, Text2Touch Key players: Allegro Hand research community, LLM+robotics labs

Two breakthroughs converge: (1) compliant in-hand rolling using vision-tactile feedback with Visiflex and TacTip sensors on Allegro Hands, and (2) LLMs autonomously designing reward functions for tactile manipulation (Text2Touch). The second is particularly notable — LLMs naturally incorporate tactile signals into reward design, suggesting they've internalized useful priors about contact-rich manipulation.

The IACAS review (Tong et al. 2024) underscores that perception systems — including tactile — remain a critical open challenge. Integrating tactile signals with whole-body humanoid control is identified as a necessary step for human-like manipulation capability.

Open problems:

• Scaling from single-primitive tasks (rolling, rotation) to multi-step manipulation sequences
• Integrating tactile policies with whole-body humanoid control
• Transferring across different sensor modalities and hand morphologies
• Reward function quality for tasks requiring fine force control

6. Enterprise Humanoid Production

Status: Rapid progress Key papers: Boston Dynamics Atlas, Tesla Optimus Gen 3 Key players: Boston Dynamics, Tesla

The enterprise humanoid market is real. Boston Dynamics' electric Atlas (56 DOF, 50kg lift, $150K, CES 2026) is targeting industrial logistics with a 30K/year factory planned for 2028. Tesla has 1,000+ Optimus Gen 3 units deployed in its own factories with a 50-100K target for 2026 and a 10M/year factory under construction. The self-deployment model (robots building robots) could create an exponential scaling flywheel.

Open problems:

• ROI demonstration for enterprise customers (2-3 year payback at $150K)
• Reliability for 24/7 factory operation
• Autonomous task adaptation vs. pre-programmed routines
• Workforce displacement and regulatory responses

7. Capability Paradigm Evolution (New from Survey Papers)

Status: Conceptual framework, tracking indicators emerging Key papers: Humanoid Robots & Humanoid AI Review, IACAS Comprehensive Review Key players: Longbing Cao (Macquarie), IACAS (CAS)

Cao (2024) introduces the most conceptually rigorous framework for evaluating humanoid progress: three paradigms (human-looking → human-like → human-level) that decouple physical appearance from cognitive capability. The "humanoid humanity dilemma" identifies the core tension: commercially polished humanoid appearance raises user expectations that current AI cannot meet, creating trust failures independent of technical progress.

The IACAS review independently converges on biomimetics and brain-inspired computing as the dual pathways for next-generation humanoid advancement — one for hardware/motion, one for cognition.

Tracking indicator: when any production humanoid begins using VLA models in real-time deployment (not just lab settings), the field will have crossed from human-looking to genuinely human-like.

Open problems:

• Standardized benchmarks for evaluating the "humanity" and "intelligence" stages
• Whether ethical/consciousness dimensions are tractable engineering problems or require AGI-level breakthroughs
• How biomimetic actuation and brain-inspired computing will be integrated into production supply chains

Recent Breakthroughs

Predictions & Trends

• Foundation models as the "brain": The pattern of VLM reasoning → task decomposition → pre-trained execution is becoming standard. VLA models will tighten this loop into end-to-end real-time control within 2-3 years.
• Teleoperation as training data pipeline: Both 1X and Figure use human operators to generate training data at scale; as autonomy improves, this bootstrapping need will decrease.
• Sim-to-real closing the gap: NVIDIA's approach of adding imperfections to simulation is more principled than domain randomization alone; 99% industrial correlation will extend to manipulation within 2-3 years.
• Consumer humanoids in 2026-2027: $20K NEO and Figure's household demos signal the market is real, even if narrow; the Cao framework suggests "human-like" capability (not just human-looking) is the gate.
• Enterprise humanoids shipping: Boston Dynamics and Tesla have moved from demos to production commitments; Tesla's self-deployment flywheel is the most consequential bet.
• Tactile sensing + LLMs converging: LLM-designed rewards for tactile policies could dramatically accelerate dexterous manipulation research; expect this to flow into humanoid hands within 2 years.
• Biomimetics as design principle: IACAS review signals that rigid-link robot design is approaching its ceiling; next-generation platforms will incorporate compliant, tendon-driven actuation.
• Two compute-demand curves, not one: training-time scaling (Xiaomi-Robotics-1's unsaturated 100k+ hours, ACE-Brain-0.5's 8B backbone) and inference-time serving (ROSA's shared-GPU-pool fleet serving) are now both concretely evidenced as separate demand curves that compound as fleets scale toward TrendForce's ~50K-unit 2026 floor — expect inference-serving infrastructure to become a distinct systems-research and compute-procurement line item over the next 12-18 months, mirroring the LLM-serving world's earlier shift from provisioning-per-request to shared multi-tenant pools.

Knowledge Gaps

Areas where the KB needs more sources:

• Physical Intelligence π*0.6 / RECAP — significant named-player gap: Physical Intelligence's RL-based self-improvement method (RL with Experience & Corrections via Advantage-conditioned Policies) reportedly more than doubles throughput on hard real-world tasks (espresso-making, laundry-folding, box assembly) versus imitation-only training. Found during this sweep but excluded as its original announcement predates the 2026-06-24 discovery cutoff (dated Nov 2025 per its model card) — flagging for an out-of-cycle ingest regardless, given it directly bears on the "humanoid foundation model recipe fragmentation" frontier (2b) and is a named player on this topic's watchlist. Suggested search/fetch: "Physical Intelligence pi-star 0.6 RECAP model card"
• Inference-serving infrastructure for robot fleets — now compiled into its own concept (Embodied-AI Inference Serving & Compute Economics) and frontier area (3b), but still only one source deep (ROSA, 2026-07-01, single preprint, not independently corroborated). This is the compute/deployment-economics layer of embodied AI (shared-GPU-pool serving, factory-objective scheduling) and connects directly to MenFem's inference-economics thesis (see cost-measurement-problem.md for the analogous LLM-side utilization argument). Suggested search: "robot fleet inference serving GPU scheduling 2026" / "vLLM robotics deployment production" / independent citations of ROSA (arXiv 2607.01088)
• VLA recipe-race entrants — independent verification pending — Xiaomi-Robotics-1, ACE-Brain-0.5, and S²-VLA (all ingested/compiled this sweep) are single preprints with author-claimed benchmarks; Xiaomi-Robotics-1's code/checkpoints are explicitly "not yet released," and none of the three has independent replication or a follow-up citing paper yet. Flagged per this sweep's evidence-discipline instruction (grade moderate, not strong, until corroborated). Suggested search: check back next cycle for code releases / citing papers for all three arXiv IDs (2607.15330, 2607.04426, 2606.27872)
• World models — additional 2026 entrants — "Orca: The World is in Your Mind" (arxiv 2606.30534) surfaced in this sweep but not ingested (cap reached); worth a look next cycle alongside the existing JEPA/generative world-model split (frontier area 4)
• Humanoid safety and human-robot interaction — suggested search: "humanoid robot safety HRI home environment 2026"
• Reinforcement learning for locomotion — suggested search: "reinforcement learning humanoid locomotion sim-to-real 2026 arxiv"
• Agility Digit deployment — suggested search: "Agility Robotics Digit deployment warehouse 2026"
• Cobot standards and regulations — suggested search: "collaborative robot safety standards ISO 2026"
• Soft robotics — not yet represented; relevant for consumer and healthcare applications with deformable bodies and compliant grippers
• Surgical robotics — high-value application domain with unique dexterity and safety requirements; zero sources in KB
• Swarm robotics — multi-robot coordination at scale; relevant for factory deployment scenarios but not yet represented
• Neuromorphic computing for robotics — IACAS review flags this as a key pathway but no dedicated sources yet
• Agility Robotics / Sanctuary AI / Apptronik — three significant humanoid companies with no KB sources
• Chinese humanoid ecosystem — partly closed (2026-06): Unitree expanded and AgiBot added as a dedicated entity (~80% Chinese-shipment duopoly). Still absent: BYD, UBTECH, Fourier, Booster Robotics as dedicated pages; the Beijing/Boston race events are ingested but their OEMs lack profiles
• Booster Robotics / Fourier — appear in deployment-cadence sources but have no dedicated entity pages despite competing in the Beijing half-marathon and the volume race