Actuate 26 | Panel Discussion | VLA vs WAM: Do Robots Need to Predict the World?

In this talk

1. 2:47

   What counts as a world model or VLA

   Danfei describes a world model as an action-conditioned predictor of future states, often expressed through video; a world action model can also predict actions intended to bring about those states. Ben defines a VLA as a low-level action policy built by adapting a vision-language model. The panel notes that the labels cover a range of architectures and training choices.

2. 5:42

   Data and deployment shape the choice

   John says off-road work and agriculture lack the location- and condition-specific driving data available to companies such as Waymo, so models that learn from video without paired actions may help. At Bonsai, a world model can generate edge cases, such as wildlife encounters, for a 3D perception system that performs collision checks. Danfei says world action models have shown advantages in dexterous manipulation, while their training and inference costs remain high.

3. 11:24

   Combine signals to fit the task

   Panelists argue that VLA and world-model approaches are not a strict either-or choice: the useful recipe depends on the robot’s task, data, and compute budget. Ben says pretrained vision-language models can contribute semantic understanding and flexible prompting, while future-state prediction may help transfer motion knowledge across human and robot embodiments. Other panelists point to tactile, force, and torque signals as useful training targets alongside video and action.

4. 35:00

   World models as simulators have limits

   A simulator can fail when a policy takes actions beyond the data used to train the model, making it difficult to keep the simulation accurate. John says imperfect models can still help Bonsai generate varied edge cases for perception training. The panel contrasts easier navigation scenarios with contact-heavy manipulation, where realistic geometry, object motion, and force or torque behavior matter.

5. 43:18

   Action-free video and compute tradeoffs

   Panelists see potential in egocentric video without action labels, but say absorbing very large video collections can be compute-intensive and annotations can improve learning within practical budgets. One example weighs a million hours of unlabeled data against 100,000 hours with action labels and 1,000 hours of robot data; the panel stresses data quality and available compute. Arman says aerospace adoption also depends on simulation fidelity, scarce task-specific data, and onboard compute.