Physical Intelligence, the San Francisco startup behind the pi0 foundation model, reached a $5.6 billion valuation after a $600 million Series B led by Alphabet Inc. The company is building what analysts call the "Android of robotics," a software layer that decouples robotic intelligence from specific hardware bodies. This approach allows a single model to control diverse machines, from industrial arms to humanoids, without rewriting code for each task. The model uses a Vision-Language-Action architecture built on Google’s PaliGemma backbone to understand semantic context, such as distinguishing a Coke can from a generic cylinder. By employing "flow matching," the system outputs continuous motor commands at 50Hz, enabling fluid movements like peeling grapes or assembling boxes. A Stanford AI Lab researcher described this progress as the "GPT-3 moment" for robotics, noting that cross-embodiment training on seven different robot types bypasses the traditional data bottleneck. You already know LLMs predict text; here, the network predicts the next physical trajectory in real time.
The Spartanburg Sequencing Test
Figure AI deployed its Figure 03 humanoid robot in Hall 52 at BMW Group Plant Spartanburg this month, moving beyond the sheet metal loading tasks performed by Figure 02 last year. The new unit handles a "sequencing use case," which involves sorting parts that do not arrive in mathematically perfect orientations within bins or carts. This task is structurally infeasible for fixed six-axis robotic arms because it requires dynamic whole-body coordination. Helix 02, Figure’s proprietary pixels-to-actions VLA, coordinates the robot’s hands, arms, torso, and feet to manipulate thin-walled parts while simultaneously pulling a heavy metal cart on caster wheels. The robot must perceive spatial variations, adjust foot placement, and shift its body weight to maintain balance during these forceful manipulations. Unlike previous deployments where Figure 02 contributed to the assembly of 30,000 cars through repetitive pick-and-place actions, Figure 03 must make small corrections on the fly as parts rotate or become partially occluded.
The commercial implications of these deployments are significant but unevenly distributed across the industry. Physical Intelligence’s open-source pi-zero checkpoints allow developers to fine-tune models using over 10,000 hours of pre-trained robot data, with inference requiring more than 8 GB of VRAM. However, the "black box" nature of learning physics through data rather than hardcoded laws creates safety concerns, as the model can exhibit unpredictable behavior in unseen scenarios. Critics argue that a robot guessing how friction works is inherently more dangerous than one following a pre-programmed safety script. While BMW assesses these robots for hazardous and repetitive tasks, the economic reality remains that only high-labour-cost environments justify the current premium pricing of humanoid systems. The gap between laboratory success rates and industrial reliability persists, leaving open whether general-purpose physical AI can scale economically before the next major hardware iteration arrives.
