Agility Robotics vs Physical Intelligence

Hardware-First vs. Intelligence-First

Agility Robotics and Physical Intelligence sit at opposite ends of the robotics value chain. Agility controls the full stack — from Digit's mechanical design and actuators to the foundation model governing its whole-body control. This vertical integration means Agility can optimize every layer for warehouse performance: Digit's 5'9" frame fits standard aisles, its 35 lb payload (50 lb in v5) matches standard tote weights, and its gait is tuned for flat warehouse floors.

Physical Intelligence takes the opposite approach: build the best possible robot brain and let others build the body. Their π0 model family accepts camera images and natural language instructions, then outputs motor commands that generalize across robot arms, grippers, and humanoid platforms. This decoupling of intelligence from hardware means Physical Intelligence can potentially capture value across the entire robotics ecosystem — but it also means they depend on hardware partners to deliver a complete product to end customers.

The question is whether robotics follows the PC model (where the OS/intelligence layer captures most of the value) or the Apple model (where vertical integration wins). History suggests both can work — but Agility is generating revenue today while Physical Intelligence is still building its platform.

Commercial Traction and Revenue Reality

Agility has the clearest commercial validation of any humanoid robot company in 2026. Digit has moved over 100,000 totes at GXO's facility. Toyota Motor Manufacturing Canada signed a RaaS agreement to deploy 7+ Digit units on its RAV4 production line after a year-long pilot. Mercado Libre is integrating Digit into Texas operations with plans for Latin American expansion. These are not pilot programs — they are commercial contracts generating revenue.

Physical Intelligence's commercial model is earlier-stage. The company has raised $1.1 billion and is valued at $5.6 billion, but revenue comes from research partnerships and licensing discussions rather than production deployments. The open-sourcing of π0 in February 2025 was a strategic move to drive adoption and build an ecosystem, but it also raises questions about monetization. If the base model is free, Physical Intelligence needs to capture value through premium model versions, fine-tuning services, or API access — a playbook familiar from the large language model world.

AI Architecture and the Foundation Model Gap

Both companies are building foundation models for robotics, but their architectures reflect their different strategies. Agility's whole-body control model is trained via sim-to-real transfer specifically for Digit's embodiment — it learns to walk while carrying objects, adjust gait for different payloads, and maintain balance during manipulation. This model is narrow but deep: it knows everything about being Digit in a warehouse.

Physical Intelligence's π0 is a vision-language-action model designed for breadth. Trained on diverse manipulation data across multiple robot embodiments, π0 can fold laundry, bus tables, assemble objects, and perform multi-step cooking tasks — often zero-shot on robot configurations it hasn't seen before. The progression from π0 to π0.5 (better open-world generalization, September 2025) to v0.6 (RECAP training that doubled throughput on complex tasks) shows rapid capability improvement. But generalization comes at a cost: π0 is unlikely to match Agility's purpose-built model on warehouse tote handling anytime soon.

The Platform Dynamics

Physical Intelligence's most intriguing strategic angle is that its customers could include Agility itself. If π0 or its successors become the default foundation model for robotic control, even hardware-focused companies might license Physical Intelligence's models rather than building their own. This mirrors the relationship between smartphone manufacturers and Google's Android — the hardware companies build the devices, but the software platform captures outsized value.

Agility's counter is that robotics is not smartphones. A warehouse humanoid needs deeply integrated hardware-software co-optimization that a general-purpose model cannot match. Digit's whole-body control model is trained specifically on Digit's dynamics, weight distribution, and actuator characteristics. Swapping in a generic model would sacrifice performance in the domain where Agility competes. The analogy is closer to Apple's custom silicon strategy: own the full stack, optimize relentlessly, and deliver a product experience that no horizontal platform can match.

Scaling and the Data Flywheel

Both companies face the robotics data challenge, but from different angles. Physical Intelligence operates teleoperation studios where human operators perform thousands of demonstrations daily, building the largest cross-embodiment robot training dataset in existence. Their explicit bet is that scaling laws apply to robotic control just as they do to language — more data, more compute, better performance. The open-sourcing of π0 also creates a community data flywheel as researchers worldwide contribute demonstrations and fine-tuning results.

Agility's data advantage is narrower but arguably more valuable per datapoint: every Digit deployed in a real warehouse generates production-grade interaction data in the exact domain Agility cares about. With units at Toyota, GXO, Amazon, and Mercado Libre, Agility is accumulating thousands of hours of real-world logistics data that no simulation or teleoperation studio can replicate. As Digit v5 scales deployments in late 2026, this data flywheel accelerates.

Safety and Regulatory Positioning

Agility is pursuing a first-mover advantage in safety certification that could become a durable competitive moat. Digit v5 targets ISO functional safety certification, which would make it the first humanoid robot cleared to work cooperatively alongside humans with no physical barriers. Current Digit models already include Category 1 stops, Safety PLCs, on-robot E-stops, and wireless teach pendants with integrated E-stops. For enterprise buyers evaluating humanoid robots, safety certification is not optional — it is a procurement requirement.

Physical Intelligence, as a software-only company, delegates safety to hardware integrators. This is both a strength (no regulatory burden) and a weakness (no control over how their models are deployed). As robotics safety regulation tightens, hardware companies with certified platforms may have an advantage over those assembling solutions from separate hardware and software providers.