Timeline

2025

January

• Jan 3 — [Paper] Gu et al. publish comprehensive survey of humanoid locomotion and manipulation — covers 3 decades of model-based approaches plus learning methods. (Survey)

April

• Apr 13 — [Paper] Wen et al. publish Humanoid-COA — first zero-shot loco-manipulation framework using foundation models (GPT-4V) on Unitree H1-2 and G1 robots. 96.6% grasping success. (Humanoid-COA)
• Apr 22 — [Paper] Physical Intelligence releases π₀.₅ — a VLA model with open-world generalization via heterogeneous co-training (multi-robot + web data + semantic prediction + low-level actions). Successor to the π₀ flow-matching VLA (10,000+ h robot-data pre-training). (π₀.₅)

June

• Jun 2 — [Paper] Humanoid World Models — open foundation world-model research targeting humanoid robotics specifically, extending the V-JEPA/tabletop and GAIA-2/AV world-model lineage to humanoid form factors. (Humanoid World Models)

October

• Oct 2025 — [Technical Report] Figure AI introduces Figure 03 — hardware and software redesign aiming for general-purpose robot that learns directly from humans. (Figure 03)

2026

January

• Jan 2026 — [Technical Report] Boston Dynamics unveils electric Atlas at CES 2026 — 56 DOF, 50kg lift capacity, 2.3m reach, estimated $150K. Shipping to Hyundai and DeepMind partners. 30K units/year factory planned for 2028. (Boston Dynamics Atlas)
• Jan 13 — [Technical Report] 1X Technologies releases world model for NEO humanoid — visual perception system enabling autonomous learning through observation. NEO preorders open at $20K. (1X World Model)
• Jan 15 — [Technical Report] Figure AI releases Helix 02 — Figure 02 demonstrates loading dishwasher, sorting packages, folding laundry using mocap data + simulation-based ML. (Helix 02)
• Jan 2026 — [Technical Report] Figure 03 appears at White House event, marking political milestone for humanoid robotics. (Figure 03)

March

• Mar 1 — [Technical Report] NVIDIA releases GR00T N1.7 — open VLA foundation model for generalist humanoid manipulation, pre-trained on ~20K hours of EgoScale egocentric human video and integrated with the Isaac Sim ecosystem; successor to N1.6, positioned for OEM licensing. (GR00T N1.7)
• Mar 9 — [Technical Report] ABB Robotics and NVIDIA announce RobotStudio HyperReality — 99% sim-to-real correlation via Omniverse integration, 0.5mm positioning accuracy. (ABB + NVIDIA)
• Mar 14 — [Paper] Ψ₀ released — open foundation model for universal humanoid loco-manipulation. Two-stage recipe (autoregressive VLM pre-training on ~800h egocentric human video → flow-based action expert on ~30h robot data) reportedly outperforms baselines trained on 10× more data. (Ψ₀)
• Mar 2026 — [Technical Report] NVIDIA Isaac Sim 5.1 adds deliberate sensor imperfections (lens defects, noise, color distortion) to force robust sim-to-real training. (ABB + NVIDIA)

April

• Apr 9 — [Technical Report] TrendForce: global humanoid shipments to breach 50,000 units in 2026 (+700% YoY). China output +94%; Unitree + AgiBot capture ~80% of Chinese shipments. Unitree commits 75K humanoid + 115K quadruped/yr; AgiBot scales 1,000 → 10,000 units. First hard volume floor for the sector. (TrendForce)
• Apr 15 — [News] Figure 03 at $39B valuation; BMW Spartanburg pilot scaling to 30k+ vehicles. Figure positions as horizontal humanoid platform vs Tesla/Hyundai vertical integration. (Figure 03 / BMW)
• Apr 19 — [News] Beijing E-Town Humanoid Half-Marathon (21.1 km). Multi-OEM humanoid endurance race — Unitree, Booster Robotics, Fourier, Agibot competing under outdoor uncontrolled conditions. Most ambitious public humanoid endurance event to date. (Beijing Half-Marathon)
• Apr 21 — [News] Tesla Optimus first US public uncontrolled debut at Boston Marathon. Cheering runners at Boylston Street showroom. Production cadence: summer 2026 low-volume, 2027 high-volume. (Optimus Boston)
• Apr 23 — [Technical Report] Princeton VLM2VLA: vision-language-action policies without catastrophic forgetting. Converts general VLMs into manipulation controllers while preserving base reasoning capability. Action plans expressed in natural language. (VLM2VLA)
• Apr 29 — [Technical Report] Figure ramps BotQ to 1 robot/hour (24x in <120 days), 350+ Figure 03 delivered; paid commercial deployment at BMW Spartanburg (~40 units). First paid commercial-scale general-purpose humanoid labor at an industrial customer. Helix VLA + new System 0 (S0) perception-conditioned whole-body controller. (Figure 03 / BotQ / BMW)

May

• May 3 — [Paper] DexSim2Real: foundation-model-guided sim-to-real for dexterous manipulation. A VLM acts as a visual-realism critic for domain randomization (FM-DR), tactile+vision fuse via cross-attention (TVCAP), and an LLM decomposes a skill curriculum (PSC). 78.2% real-world success across six contact-rich tasks; sim-to-real gap cut to 8.3%, beating DrEureka and DeXtreme. (DexSim2Real)

June

• Jun 1 — [Industry] Unitree's Shanghai IPO cleared by the SSE listing committee — first "embodied AI" / humanoid pure-play approved for China's A-share market. ~$6.2B target valuation, ~$616M (4.2B yuan) raise. (context: TrendForce)
• Jun 5 — [Paper] The Sim-to-Real Gap of Foundation Model Agents (KDD 2026 Blue Sky): reframes LLM-agent robustness as a classical sim-to-real problem formalized over the four MDP elements (Observation/Action/Transition/Reward); advocates importing robotics' domain randomization. Inverts the usual ML→robotics flow of ideas. (MDP Agent Sim2Real)
• Jun 26 — [Paper] S²-VLA: compact 2B-parameter VLA beats 7B-scale baselines on long-horizon manipulation (LIBERO, SimplerEnv) via State-Space Guided Adaptive Attention — a belief-state tracker over task progression that dynamically gates visual/language/action fusion. Architecture-over-scale result. (S²-VLA)
• Jun 30 — [Paper] First sim-to-real transfer of a world-action model: Cosmos Policy (video-diffusion) trained on ~800 purely-synthetic demonstrations per task with zero real-world data reaches 35% average zero-shot success on a physical Franka arm across lifting, drawer-opening, and pick-and-place. (Sim-to-Real World-Action Models)

July

• Jul 1 — [Paper] ROSA: robotics foundation-model serving system for robot factories (Stanford/NVIDIA). Shared GPU-pool serving replaces the per-robot/edge-GPU assumption; factory-objective scheduling on Ray Serve/vLLM/PyTorch/JAX reports up to 12.06x factory-productivity improvement over dedicated serving. First systems paper to treat embodied-AI inference serving as a distinct compute-economics problem. (ROSA)
• Jul 5 — [Paper] ACE-Brain-0.5: unified 8B-parameter embodied foundation model organizing robot intelligence into five coupled functions (spatial perception, decision-making, embodied interaction, self-monitoring, self-improvement) via a new SSR+ training recipe; improves 14/18 spatial-perception benchmarks over ACE-Brain-0. (ACE-Brain-0.5)
• Jul 16 — [Paper] Xiaomi-Robotics-1: VLA pre-trained on 100k+ hours of UMI-collected real-world trajectories with an auto-labeling pipeline; scaling reported as unsaturated across data volume and model size; new SOTA on RoboCasa365 (57.4% vs. 46.6%) and RoboDojo (20.07 vs. 13.07). First major corporate (non-hardware-vendor, non-lab) entrant into the VLA foundation-model race. (Xiaomi-Robotics-1)

Q1–Q2

• 2026 — [Technical Report] Tesla deploys 1,000+ Optimus Gen 3 units across its factories for autonomous parts processing. 50,000-100,000 unit production target for 2026. (Tesla Optimus)
• Q2 2026 — [Technical Report] Tesla retires Model S/X lines at Fremont to convert floor space for Optimus (line designed for ~1M units/yr); Optimus V3 production confirmed for summer 2026 (high volume summer 2027). Dedicated 10M units/year Giga Texas factory under construction. (Tesla Optimus)
• 2026 — [Paper] Compliant in-hand rolling manipulation demonstrated using tactile feedback on Allegro Hand with Visiflex fingertips. (Tactile Rolling)
• 2026 — [Paper] Text2Touch: LLMs autonomously design reward functions for tactile manipulation — matches hand-engineered rewards for in-hand rotation on real TacTip hardware. (Text2Touch)

Q2-Q4 (Upcoming/In Progress)

• Q2 2026 — [Catalyst] Boston Dynamics Atlas fleets shipping to Hyundai RMAC (first customer) and Google DeepMind. All 2026 production committed. Tasks: component sequencing, machine tending, order fulfillment. 56 DOF, 50kg lift, 2.3m reach. Additional customers onboarding early 2027.
• Q3-Q4 2026 — [Catalyst] 1X NEO first consumer deliveries (US). $20K Early Access purchase or $499/month subscription. 66 lbs (30 kg), lifts 150 lbs, 22 DOF hands, 22 dB noise. Note: Not fully autonomous in 2026 — requires teleoperator support for complex tasks. Developing platform, not finished product.
• 2026 — [Industry] Hyundai announces new robotics factory capable of producing 30,000 robots/year for Atlas scale-up.