PsiBot Raises $280M for Embodied AI

🔑 Enhanced Key Takeaways

• •PsiBot was founded on September 1, 2024, by Dr. Viktor Wang, who has nearly two decades of experience in mobile devices, smart speakers, and robotics, with co-founder Yuanpei Chen, a post-2000s robotics prodigy trained at Stanford under Fei-Fei Li and Karen Liu[1][2][5].
• •The company developed custom exoskeleton devices for 1:1 real-world physical data collection, enabling supervised fine-tuning (SFT) and real-world reinforcement learning to optimize model performance beyond pretraining[3].
• •PsiBot was the only Chinese embodied AI company invited to the 2025 Global Developer Pioneer Conference, where it unveiled its VLA foundation model and presented breakthroughs in multimodal penetration and long-horizon skill chaining[3].
• •R1 model demonstrated 30-minute sustained reasoning and manipulation by playing live Mahjong with humans, using 'Chain of Action Thought' (CoAT) integrating perception, reasoning, and execution via an Action Tokenizer[5].

🛠️ Technical Deep Dive

• •Hierarchical end-to-end RL framework: Psi-P0 planning model for task decomposition, reasoning, and generalization using large models to understand action-environment interactions; Psi-C0 control model for execution[1][2].
• •Psi-R0, R0.5, R1: Industry-first end-to-end RL embodied models achieving long-horizon tasks with dual-handed dexterous manipulation, strong generalization across objects/scenarios, and full perception-to-action loops[1][3][5].
• •Data pipeline: Custom exoskeletons for real-world data collection, SFT fine-tuning, followed by real-world RL optimization; supports lifelong learning via memory structures[2][3].
• •R1 architecture: Separates planning/control connected by Action Tokenizer; enables Chain of Action Thought (CoAT) for extended tasks like 30-minute Mahjong with strategic reasoning[5].
• •Psi-P0 surpasses OpenAI's VPT and Nvidia's MineDojo in task complexity and accuracy for open-world planning[2].

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline