Dexterity launches Foresight world model and 4D packing agent

Dexterity, a provider of physical AI and robotics, said it has taken a major leap forward in its physical AI stack, anchored by Foresight, its new world model and 4D box packing agent.

The company said that these advancements help solve some of the most physically demanding and hardest-to-staff tasks, such as truck loading.

Alongside the announcement, Dexterity also launched the Foresight API Challenge with up to $50,000 in prizes for student teams.

A world model built for the physical world

Dexterity said that ForeSight is a physics-consistent world model, generating a real-time, transactable representation of the physical environment that enables robots to perceive, reason and act. The company said that Foresight represents a new class of world model, built not for observation, but for physical manipulation at the production scale.

In autonomous truck loading, Foresight powers Dexterity's Mech dual-armed “superhumanoid” robot, with a 4D box packing agent that reasons across three spatial dimensions plus time, determining where to place each package onto an evolving wall of freight.

Dexterity said that 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, jointly optimizing density, stability, reachability and dual-arm parallelism, while predicting how each placement affects the integrity of the entire truck.

An interpretable, safety-first architecture

Built on Foresight, Dexterity's agentic framework coordinates perception, decision and motion agents that operate asynchronously to automate truck loading, package sortation and other applications. Dexterity said that the architecture is interpretable and safety-first, giving operators visibility into why the system makes each decision.

"Foresight delivers real-time, production-grade random box packing in 4D space-time, predicting how one placement dictates the integrity of the entire truck," said 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 said that this Physical AI stack is application-agnostic and hardware-agnostic: it is proven in production across six applications and a developer platform, running on four robot types and five hand types. To date, Dexterity said that Foresight has been trained with experience from over 100 million autonomous actions in production.

Foresight API Challenge

To give the Physical AI community a window into production-grade world models, Dexterity also announced the launch of the Foresight API Challenge in March.

Student teams can build packing agents and compete on a public leaderboard for up to $50,000 in prizes. Dexterity said that no simulator is provided, and competitors must build their own understanding of the physics.

Dexterity also launched a browser-based truck loading game that lets anyone experience the problem firsthand.