Danijar Hafner is developing AI agents that can plan for unexpected situations by navigating environments they have not previously encountered. He achieves this using model-based reinforcement learning to create world models that simulate physical reality for the agents. This technique allows agents to make predictions about future outcomes, enabling them to operate in unfamiliar physical spaces like a person's home. Previous work, like Dreamer, demonstrated agents solving complex tasks without extensive real-world trial-and-error training. Hafner is now migrating this approach to physical robots to allow them to operate in novel environments.
XPENG has initiated mass production of its IRON humanoid robot, claiming a global first for assembly and autonomous operation without human input. The production line utilizes automotive-grade quality control and over 80% automation, leveraging an 85% supply chain overlap with XPENG's EV infrastructure. The robot features three in-house Turing AI chips and capabilities like multi-modal dialogue and autonomous navigation. Initial units are being deployed in showrooms and campuses before broader enterprise deliveries are slated for 2027. This move positions XPENG to challenge Tesla in the race for scaled humanoid robotics manufacturing.
Chinese financial outlets report Tesla has placed initial multi-thousand-unit component orders for the Optimus program across core suppliers. This procurement follows reported line commissioning at the Fremont facility and aligns with Tesla's internal goal to deliver 10,000 to 20,000 units by 2026. These orders caused immediate market gains for domestic robotics suppliers specializing in actuators and vision hardware. The move is viewed as a transition of the Optimus initiative from prototyping to a structured, production-level supply chain ramp. This development signals a focus on securing the critical mechanical components necessary for mass humanoid robot production.
Unitree Robotics demonstrated the UnifoLM-X2-1.0 world model, enabling fully autonomous sparring for its G1 humanoid robot. This system uses predictive trajectory rollouts to anticipate opponent movements and execute dynamic strikes without human teleoperation. The world model allows the robot to predict future scene changes and opponent motion to select optimal motor trajectories in high-speed interactions. The implementation utilizes a client-server pipeline where heavy inference runs offboard, addressing latency bottlenecks in world-action models. This advancement aims to prove the feasibility of large-scale, world-model-driven physical AI deployment in robotics.