Humanoid Physical AI Platforms

Nvidia is the default substrate for humanoid Physical AI

Nvidia leads the investor table with 50 deals, ahead of Blackstone (8), AMD Ventures (5), BlackRock and Coatue (4 each). Its role goes well beyond chips: GR00T N1 and N1.7-3B are the reference open humanoid foundation models, and papers run whole-body controllers at 50Hz on RTX 5090s and train in GPU-accelerated MuJoCo with 4,096 parallel environments. Nvidia also plans to invest in Genesis AI, and has moved into models (Nemotron, a $30B Reflection acquisition). The stack is vertically integrated, so humanoid builders increasingly compete and co-develop on Nvidia's terms.

Weekly funding peaked at $10.85B (week of 7/27) and $5.5B on a single deal (week of 9/7). It has since fallen to $513M, $210M and $72M in the latest three weeks, even as mentions hold steady at 73-94 a week. Seed alone absorbed $5.4B across 7 deals, and Genesis AI's $105M seed co-led by Eclipse and Khosla is an example of the foundation-model-first, large early-stage round. Strategic corporates and private credit, including Blackstone, are increasingly the funders.

Why it matters · Attention remains high while checks shrink, so late entrants face a market where the leaders have already raised and the rest compete for much smaller rounds.

Physical Intelligence's π0 and π0.5 are the default base policy for real-world manipulation experiments, including on DROID, and Nvidia's GR00T is the humanoid-specific counterpart. New research directly challenges their continuous-action design: discrete action parametrizations match or beat continuous ones across imitation learning, multi-task VLA training and real-world tasks. Robustness varies much more than headline success. LIBERO-Plus scores range from 73% to 79% across encoders, even though standard LIBERO results sit near 95%.

Why it matters · The policy layer is not yet settled, so architecture shifts could erode the moat of any single VLA vendor.

Diffusion-generated humanoid policies reach 1.000 best-of-100 success on walking, dancing and spin kick and deploy zero-shot on a real Unitree G1 with no fine-tuning. Distilled whole-body policies also transfer without privileged object state. Classical symbolic planning is clearly weaker: the TAMP system TiPToP scores 9.4% against 11.0% for the raw policy. Researchers also argue the next unlock is whole-body contact, rather than the end-effector dexterity emphasized by Figure, Boston Dynamics and Apptronik.

Why it matters · Neural whole-body control makes sim-trained behavior cheap to port across hardware, shifting differentiation from teleoperated data to simulation and policy distillation.

Unitree's G1 was the sole real-world platform in multiple recent papers, and Dyna's CEO called its small sim-to-real gap a strong breakthrough. This pairs low-cost Chinese hardware with Western foundation models and Nvidia compute. Chinese universities (Tsinghua, Peking, Shanghai Jiao Tong) are also contributing embodied-AI research.

Why it matters · A cheap shared hardware baseline lowers entry barriers for model startups and puts pricing pressure on full-stack humanoid makers.