Physical AI: A 4B Vision-Language-Action Model and the Sim-to-Real Gap We Do Not Hide
Robots need a different kind of model
A chatbot's output is text. A robot's output is torque, grasp, and motion - actions that fail in the physical world if the model misjudges geometry, timing, or contact. Physical AI is VIDRAFT's research line for that harder output space: vision-language-action (VLA) models that take pixels and instructions in, and emit actions out.
Darwin-4B-Robo
Darwin-4B-Robo is our VLA model. On the RoboCasa 24-task benchmark it ranked #1 (leaderboard scope) - a simulated manipulation suite covering 24 distinct household-style tasks. At 4B parameters it is deliberately small: embodied policies have to run on real robots with real latency budgets, so parameter efficiency is a feature, not a compromise.
The model builds on the David + pi-0 lineage - prior open work in the VLA space - rather than starting from scratch. Standing on that lineage is the honest way to move fast in robotics research, where reproducible baselines are scarce.
Why small models, and why simulation first
Embodied research runs on a tight loop: a policy that cannot execute inside a robot's control cycle is useless no matter how accurate. That is why Darwin-4B-Robo is small. Simulation benchmarks like RoboCasa let us iterate hundreds of times faster than on hardware and compare against other methods on equal footing - but we treat sim numbers as a leading indicator, not a finish line. The real test is always physical, and that is the harder, slower work ahead.
World models
The longer-term thread is world-model research: giving a policy an internal predictive model of how the environment responds to its actions, so it can plan rather than react. A VLA that only maps observation to action is reactive; one that can roll out "if I push here, this happens" is a step toward robust manipulation. This part is early.
Honest scope
The RoboCasa #1 result is scoped to that benchmark's leaderboard and its task set; RoboCasa is a simulation suite, so it measures simulated manipulation, and sim performance does not automatically transfer to physical robots (the sim-to-real gap is real and unsolved in general). Darwin-4B-Robo is embodied-AI research, not a shipping robotics product, and the world-model direction is research-stage. The David + pi-0 lineage is prior work we build on and credit, not our invention. Read this as a research report from an active line, not a product datasheet.
More: https://vidraft.net
Top comments (0)