Figure AI vs Physical Intelligence

Vertical Integration vs. Platform Play

The most consequential difference between Figure AI and Physical Intelligence is where each company draws its system boundary. Figure builds the entire robot — chassis, actuators, sensors, hands, and the AI that controls them — optimizing every layer of the stack together. Physical Intelligence builds only the intelligence layer, designing it to be robot-agnostic from the start.

Figure's approach mirrors Apple's playbook: tight hardware-software integration enables capabilities that general-purpose solutions struggle to match. Their Helix system's dual VLM/VLA architecture is co-designed with Figure 03's sensor suite, from the palm-embedded cameras to the tactile sensors that detect forces as small as three grams. Physical Intelligence's approach mirrors Android or, more precisely, the way OpenAI's GPT models became a general-purpose reasoning layer adopted across thousands of applications. If π0 becomes the default foundation model for robotic control, Physical Intelligence captures value across the entire ecosystem.

The risk for Figure is building hardware in a market where Chinese manufacturers like Unitree already ship roughly 36 times more units. The risk for Physical Intelligence is that robot intelligence may be too tightly coupled to specific hardware to truly generalize — that the "write once, run anywhere" promise breaks down when physics is involved.

AI Architecture: Dual-System vs. Unified Model

Figure's Helix uses a deliberately split architecture: a vision-language model handles scene understanding and task reasoning, while a separate vision-language-action model generates motor commands at 200Hz. This separation lets Figure update its reasoning capabilities without retraining its motor control, and vice versa — a practical engineering advantage when deploying to production robots that need both intelligence and real-time responsiveness.

Physical Intelligence's π0 takes the opposite bet: a single unified vision-language-action model that processes camera input, language instructions, and action generation in one forward pass. The flow-matching architecture lets π0 generate smooth, continuous action trajectories rather than discrete motor commands. Their March 2026 work on Multi-Scale Embodied Memory (MEM) adds both long-term and short-term memory to the model, enabling complex tasks exceeding ten minutes — addressing a key limitation of earlier VLA approaches.

Both architectures have demonstrated impressive results. Helix 02 completed a four-minute autonomous dishwasher task with no resets. π0 has demonstrated zero-shot generalization across robot platforms it was never explicitly trained on. The question is whether the future favors specialized, tightly-coupled systems or general-purpose models that improve with scale.

Data Strategy and the Scaling Hypothesis

Both companies are betting heavily on the foundation model scaling hypothesis — the idea that more data and compute will yield increasingly capable robot intelligence — but they are collecting data in very different ways.

Figure operates its own teleoperation fleet where human operators remote-control Figure 02 and Figure 03 robots to generate manipulation data at scale. This data is proprietary, specific to Figure's hardware, and tightly quality-controlled. Physical Intelligence collects data across multiple robot embodiments from multiple partners, combines it with internet-scale pre-training, and has open-sourced its model to encourage community contributions via the openpi repository on GitHub.

Physical Intelligence's cross-embodiment approach potentially gives it access to a much larger and more diverse training set. Their RECAP training pipeline — which combines demonstration, coaching through corrections, and autonomous experience — doubled throughput on precision tasks in their v0.6 release. Figure's single-embodiment data may be narrower but deeper, with every demonstration perfectly matched to the robot that will execute the task.

Commercial Traction and Go-to-Market

Figure AI has the more visible commercial deployments. Their 11-month pilot at BMW's Spartanburg plant — loading 90,000+ automotive parts and assisting production of 30,000+ BMW X3 vehicles — is one of the most substantial humanoid robot deployments in industry history. The Figure 03, designed for home use at a target price of $20,000, signals Figure's ambition to move beyond industrial settings, though consumer availability is not expected until late 2026 at earliest.

Physical Intelligence's commercial traction is earlier-stage but potentially broader. Partner Weave has robots performing live deployments in a San Francisco laundromat. AgiBot has deployed Physical Intelligence's reinforcement learning system in a manufacturing pilot. Because Physical Intelligence doesn't need to manufacture or ship hardware, it can scale deployment through software updates to existing partner robots — a fundamentally different growth curve.

The BotQ manufacturing facility gives Figure a concrete production roadmap: 12,000 humanoids per year initially, scaling to 100,000 over four years. Physical Intelligence's scaling is measured in model downloads and partner integrations rather than units shipped.

Open Source and Ecosystem Strategy

Physical Intelligence made a significant strategic move by open-sourcing its π0 and π0.5 model weights and fine-tuning code. This mirrors the playbook of companies like Meta with Llama — building ecosystem adoption and developer familiarity that can later be monetized through enterprise offerings and premium model access. The openpi repository includes checkpoints for multiple robot platforms, example inference code, and fine-tuning pipelines.

Figure AI keeps its Helix system entirely proprietary. This makes sense for a vertically integrated hardware company — the AI is the differentiation that justifies the hardware price — but it limits Figure's ability to build a broader ecosystem. If the robotics market evolves like the smartphone market, the question is whether Figure is building the iPhone (premium, integrated, profitable) or whether Physical Intelligence is building Android (ubiquitous, platform-dominant, ecosystem-defining).

Funding, Valuation, and Investor Thesis

Figure AI's $39 billion valuation is roughly 7x Physical Intelligence's ~$5.6 billion valuation, reflecting investor conviction in the vertically integrated approach and the sheer scale of the addressable market for humanoid robots. Figure's backers — Microsoft, NVIDIA, OpenAI, Intel, and Jeff Bezos — read like a who's-who of the agentic AI ecosystem.

Physical Intelligence's funding trajectory is remarkable in its own right: from founding in 2024 to $1.1 billion raised in under two years. The platform thesis — that Physical Intelligence could become the "intelligence layer" for all robots — justifies a venture-scale bet even at a lower valuation. If the robotics market fragments across many hardware makers (as seems likely given Chinese manufacturing advantages), the software platform that powers them all could be more valuable than any single hardware company.