Dexterity Inc. today announced advancements in its physical AI stack for industrial robots, anchored by Foresight, a world model and 4D box-packing agent. The company said these advances can help with some of the most physically demanding and hardest-to-staff tasks, such as truck loading.
In addition, Dexterity is launching the Foresight API (Application Programming Interface) Challenge with up to $50,000 in prizes for student teams.
Founded in 2017 at Stanford University’s robotics lab, Dexterity said it “builds full-stack physical AI systems to address the most labor-intensive tasks in warehouse and logistics operations.” The Redwood City, Calif.-based company described Foresight as “a physics-consistent world model, an agentic skill framework, and interpretable safety-first architecture.”
Leading enterprises around the world use Foresight for truck loading and unloading, parcel singulation, and palletizing and depalletizing, added Dexterity.
Foresight models the physical world
Foresight is a transactable representation of the physical environment that enables robots to perceive, reason, and act. “Foresight represents a new class of world model, built not for observation, but for physical manipulation at the production scale,” asserted Dexterity.
In autonomous truck loading, Foresight powers Mech, Dexterity’s dual-armed “superhumanoid” robot. The 4D box-packing agent reasons across three spatial dimensions plus time, determining where to place each package onto an evolving wall of freight.
This is a combinatorial problem far more complex than the game of Go, with near-infinite input variation, up to 400 potential placements per box, and multiple walls packed simultaneously. Foresight makes each placement decision in under 400 milliseconds, explained Dexterity.
The system optimizes density, stability, reachability, and dual-arm parallelism while also predicting how each placement affects the integrity of the entire truck, the company said.
Architecture is interpretable, safety-first
Built on Foresight, Dexterity claimed that its agentic framework coordinates perception, decision, and motion agents that operate asynchronously to automate truck loading, package sortation, and other applications.
“The architecture is interpretable and safety-first, giving operators visibility into why the system makes each decision,” noted the company.
Dexterity added that its physical AI stack is application-agnostic and hardware-agnostic, proven in production across six applications and a developer platform. It is running on four robot types and five hand types. To date, Foresight has been trained with experience from over 100 million autonomous actions in production.
“Foresight delivers real-time, production-grade random box packing in 4D space-time, predicting how one placement dictates the integrity of the entire truck,” stated Samir Menon, founder and CEO of Dexterity. “Physical AI is not just a future promise; it is a system that perceives, decides, and acts in the real world, right now.”
Dexterity launches Foresight API Challenge
To give the physical AI community exposure to production-grade world models, Dexterity has launched the Foresight API Challenge. Student teams build packing agents and compete on a public leaderboard for up to $50,000 in prizes.
No simulator is provided; competitors must build their own understanding of the physics. Challenge details and signups are available at dexterity.ai/challenge.
In addition, Dexterity said its browser-based truck-loading game lets anyone experience the problem firsthand. The company was a 2024 RBR50 Robotics Innovation Award honoree for its trailer-unloading system.