Physical AI & Robotics Race
Key Questions
What funding trends are emerging in physical AI and robotics?
Robotics VC hit a record $55.8B mid-2026, driven by software-defined VLA models. Skild reached $14B valuation as AI infrastructure, with Neura Robotics at $1.4B and Nvidia GR00T positioning as a robotics platform.
Which new models and benchmarks advance robotic capabilities?
NVIDIA Cosmos 3 serves as an open foundation model, while RoboTTT enables context scaling for 87% gains and one-shot imitation from video. RoboDojo benchmark covers 42 sim and 18 real manipulation tasks.
When might robotics reach a 'ChatGPT moment' according to industry leaders?
Zhiyuan Robotics projects a potential breakthrough in 2-5 years after overcoming data, representation, and closed-loop walls. The company targets 99.99% success rates on 3C lines using 100M hours of human data.
Robotics VC breaks annual record at $55.8B mid-2026, driven by shift to software-defined VLA models. Skild valued at $14B as 'AI infrastructure'; Nvidia GR00T positions as Android-for-robotics. ENPIRE enables autonomous physical-world research. Neura Robotics $1.4B, Theker €73M. NVIDIA Cosmos 3 open frontier foundation model. Flash-WAM distillation 23x speedup. VisualClaw real-time physical world agents. DreamX-World 1.0 interactive world model. PAIWorld for robotic manipulation. Manifold AI raises funds for world models. Collecting robot training data is dirty work. New paper: WorldDirector decouples semantic motion from visual generation using LLM-coordinated 3D trajectories with persistent dynamic memory, advancing controllable world simulators for robotics and content creation. Embodied.cpp portable inference runtime for embodied AI on heterogeneous robots, unifying VLA and WAM models with high task success rates and memory reduction. GigaWorld-1 roadmap for world models for robot policy evaluation introduces WMBench and GigaWorld-1, emphasizing long-horizon consistency over visual realism. New today: RoboDojo benchmark for generalist robot manipulation (42 sim, 18 real tasks); LingBot-Video open-source MoE video foundation model for embodied intelligence. Also today: RoboTTT context scaling for robot policies (8K timesteps, 87% gain over single-step baseline) enables one-shot imitation from human video and on-the-fly improvement, advancing robot foundation models. New today: Zhiyuan Robotics projects ChatGPT moment for robots in 2-5 years, identifying three walls (data, representation, closed-loop) and targeting 99.99% success rate on 3C lines with 100M hours of human data.