Physical AI

Physical AI data comes from cameras, lidar, millimeter-wave radar, IMUs and other sensors, in various formats and at massive scale.

Open models in the physical AI field allow small and medium-sized research teams and startups to fine-tune based on existing weights and architectures without having to build foundation models from scratch.

Training for physical AI requires large amounts of sensor data, simulation environments, and hardware-in-the-loop testing, with a resource threshold higher than that of cloud AI.

The current bottleneck in physical AI is concentrated in data rather than compute: interaction data with physical annotations is extremely scarce.

Physical systems are extremely sensitive to latency, and inference must be able to run efficiently on edge devices, imposing requirements on model compression and hardware co-design that differ from those of cloud AI.

Physical AI is far more fragmented than language, and there may not be a few unified foundation models; instead, several open models for specific domains will coexist.

Physical AI refers to the technological direction of taking AI beyond screens and into the real world to interact with the physical environment, with typical deployment scenarios including robots, automated equipment, and edge sensing systems.

Embodied reasoning refers to the capability of an AI model, given a physical body, to perceive, understand, and plan within its environment and execute actions.

The difficulty of reproducing physical AI experiments includes environmental uncontrollability, as physical variables such as temperature, lighting, and ground friction coefficients are difficult to precisely replicate across different laboratories.

Reproducing experiments in physical AI (robot manipulation, embodied intelligence, etc.) often requires expensive hardware equipment and even laboratory environments equipped with high-speed cameras.