Dexterity World Model Speeds Up Robot Decisions on NVIDIA Hardware
- Dexterity’s Foresight is a world model that builds a real-time view of physical environments, allowing robots to evaluate geometry, contact, and stability before taking action.
- Running on NVIDIA hardware, the system’s visual perception cycle dropped from 1,508 milliseconds to 90 milliseconds, delivering a 17x speedup across its core pipeline.
Dexterity says its Foresight system, a world model designed for Physical AI, powers how its robots understand physical environments before they act. The model builds a real-time view of the world from sensor data and uses it to plan actions, accounting for factors such as object geometry, contact dynamics, weight distribution, and stability.
Source: Dexterity
Running on NVIDIA hardware, Dexterity focused on speeding up how its world model is built from sensor data. The company says it reduced the time it takes to process each scene from 1,508 milliseconds to 90 milliseconds, a 17x improvement, by optimizing three steps: identifying objects, building their 3D structure, and applying physics rules to understand how they interact. This included using TensorRT, parallel GPU processing, and custom CUDA code, while also allowing the system to use full-resolution sensor data instead of a reduced sample.
Dexterity frames this as a move toward real-time use in production. The faster system can run dual-arm robots continuously, increasing throughput while cutting down on placement errors in truck loading. It showcased that work at FedEx’s 2026 Investor Day, where its Mech robot autonomously loaded a trailer with randomly shaped packages, and FedEx said it intends to scale trailer loading and unloading across several U.S. hubs over the next few years.
🌀 Tom’s Take:
Faster perception means the robot can act on a current view of the world instead of an outdated one, which is critical when handling many objects in real time.
Sources: PR Newswire / Dexterity
