BAIR Introduces GRASP: Gradient-Based Planning for Long-Horizon World Models
Researchers from the Berkeley Artificial Intelligence Research (BAIR) lab have introduced GRASP, a novel gradient-based planner designed to enhance the utility of learned world models for long-horizon planning. While modern world models are becoming powerful general-purpose simulators capable of predicting high-dimensional visual sequences, using them for effective control remains challenging due to optimization fragility, ill-conditioning, and bad local minima. GRASP addresses these issues through three key innovations: lifting trajectories into virtual states to enable parallel optimization across time, injecting stochasticity directly into state iterates to improve exploration, and reshaping gradients to provide clean signals for actions while avoiding brittle state-input gradients in high-dimensional vision models. This approach aims to make gradient-based planning more robust and practical for complex tasks. The work, collaborated on by experts including Yann LeCun, highlights significant advancements in making differentiable simulators more effective for robotics and AI control systems, moving beyond simple short-horizon predictions to reliable long-term strategic planning.
Wire timeline
BAIR Introduces GRASP: Gradient-Based Planning for Long-Horizon World Models
Researchers from the Berkeley Artificial Intelligence Research (BAIR) lab have introduced GRASP, a novel gradient-based planner designed to enhance the utility of learned world models for long-horizon planning. While modern world models are becoming powerful general-purpose simulators capable of predicting high-dimensional visual sequences, using them for effective control remains challenging due to optimization fragility, ill-conditioning, and bad local minima. GRASP addresses these issues through three key innovations: lifting trajectories into virtual states to enable parallel optimization across time, injecting stochasticity directly into state iterates to improve exploration, and reshaping gradients to provide clean signals for actions while avoiding brittle state-input gradients in high-dimensional vision models. This approach aims to make gradient-based planning more robust and practical for complex tasks. The work, collaborated on by experts including Yann LeCun, highlights significant advancements in making differentiable simulators more effective for robotics and AI control systems, moving beyond simple short-horizon predictions to reliable long-term strategic planning.
The Berkeley Artificial Intelligence Research Blog