WAM
A robot control paradigm that performs planning and decision-making by learning an internal world model of how the environment changes in response to actions. It emphasizes modeling the causal relationships of the physical world to achieve stronger generalization capabilities and few-shot learning.
Core Facts
Timeline (last 90 days)
The key difference between VLA and WAM compared with general MLLMs is that they have undergone action-space alignment, with outputs falling directly within the robot's controllable instruction space.
The Harness+policy mode still falls below the per-task performance ceiling of each component running alone, for reasons including suboptimal policy invocation, lack of subgoal decomposition in fixed instruction execution, and interference with grasping when the agent regains control.
When the agent is allowed to invoke post-trained VLA or WAM policies, overall task success rate rises to the level of the strongest comparison policy, while token cost drops significantly.
Jia Xiaosong observed that models such as VLA and WAM that introduce stronger vision-language capabilities are not necessarily better than well-optimized end-to-end models on some autonomous driving benchmarks; what truly limits the model is whether it can make the right decision, not whether it recognizes a certain object.
VLA leans toward end-to-end imitation and builds capability through data scale, while WAM leans toward explicit world modeling and improves generalization through structured reasoning.
WAM proponents argue that once predictive models are sufficiently accurate, agents can achieve few-shot or even zero-shot motion planning through model predictive control (MPC) or latent-space tree search, bypassing the reliance on large volumes of real demonstration data.
LIBERO simulation successfully runs long-horizon tasks on Mac for the first time, with GR00T achieving a 100% success rate, WAM achieving 98%, and a 2.8-billion-parameter VLA small model achieving 96%.
World Action Models (WAMs) couple predictive world modeling with action generation, enabling the model not only to predict the next state of the environment but also to guide agents in taking their next actions accordingly.
All Facts (8)
World Action Models (WAMs) couple predictive world modeling with action generation, enabling the model not only to predict the next state of the environment but also to guide agents in taking their next actions accordingly.
75%UnverifiedThe Harness+policy mode still falls below the per-task performance ceiling of each component running alone, for reasons including suboptimal policy invocation, lack of subgoal decomposition in fixed instruction execution, and interference with grasping when the agent regains control.
50%UnverifiedThe key difference between VLA and WAM compared with general MLLMs is that they have undergone action-space alignment, with outputs falling directly within the robot's controllable instruction space.
50%UnverifiedWhen the agent is allowed to invoke post-trained VLA or WAM policies, overall task success rate rises to the level of the strongest comparison policy, while token cost drops significantly.
50%UnverifiedJia Xiaosong observed that models such as VLA and WAM that introduce stronger vision-language capabilities are not necessarily better than well-optimized end-to-end models on some autonomous driving benchmarks; what truly limits the model is whether it can make the right decision, not whether it recognizes a certain object.
50%UnverifiedWAM proponents argue that once predictive models are sufficiently accurate, agents can achieve few-shot or even zero-shot motion planning through model predictive control (MPC) or latent-space tree search, bypassing the reliance on large volumes of real demonstration data.
50%UnverifiedVLA leans toward end-to-end imitation and builds capability through data scale, while WAM leans toward explicit world modeling and improves generalization through structured reasoning.
50%UnverifiedLIBERO simulation successfully runs long-horizon tasks on Mac for the first time, with GR00T achieving a 100% success rate, WAM achieving 98%, and a 2.8-billion-parameter VLA small model achieving 96%.
50%