Oct 13, 2026 · 1h 12m
Dyna Robotics pushes physical agents beyond impressive demos
Dyna-2: General Purpose Physical Agents with Hour-Long Autonomy — Jason Ma, Dyna Robotics
Reliable commercial robots require advances across data, hardware, inference, and evaluation rather than a single breakthrough model.
- 1Dyna-2 combines large-scale human and robot data with world-state prediction and executable action generation.
- 2Dyna keeps inference on-device and controls much of the hardware-software stack to meet robotics’ stringent latency demands.
- 3Commercial deployments show general-purpose models handling deformable objects while exposing the unresolved challenge of long-horizon autonomy.
Don't miss
Jason describes Dyna’s deployments folding laundry at Best Western and large volumes of napkins at Din Tai Fung, showing a general-purpose model moving into production work.
The brief
Jason Ma presents Dyna-2 as a world-action model trained on more than one million hours of human video and robot data, predicting future states while generating executable actions.
The strategy pairs a general-purpose base model with fast, on-device inference and a tightly integrated hardware-software stack, reflecting how physical agents face tighter latency constraints than software agents.
Dyna’s semi-humanoid, wheeled platform and teleoperated task adaptation are designed for practical manipulation, while synthetic environments and multimodal sensing may be needed to overcome scarce robotics data.
Deployments folding laundry at Best Western and napkins at Din Tai Fung mark a shift from polished demonstrations toward measurable production work with deformable objects.
The central uncertainty is long-horizon autonomy: scaling robots across industries will require faster physical evaluation, richer sensing, stronger hardware, and much more diverse real-world data.