Gradient-Based Planning for Long-Horizon World Models
GRASP makes long-horizon planning practical with learned dynamics.
1What is GRASP?
GRASP is a gradient-based planner for learned dynamics that makes long-horizon planning practical by lifting trajectories into virtual states, adding stochasticity, and reshaping gradients.
2World Models
World models are learned models that predict future outcomes given the current state and future actions. These models can predict long sequences in high-dimensional visual spaces.
3Planning Process
Planning involves optimizing action sequences through the model to minimize terminal error, given a start state and a goal.
4Challenges of Long-Horizon Planning
Long-horizon planning is challenging due to deep and ill-conditioned computation graphs, landscapes full of traps, and issues specific to learned models.
5GRASP's Solutions
GRASP addresses long-horizon planning challenges by lifting into virtual states, adding stochasticity, and reshaping gradients for robustness.
6Lifting into Virtual States
GRASP lifts trajectories into virtual states to ensure optimization is parallel across time.
7Adding Stochasticity
GRASP adds stochasticity directly to state iterates for exploration.
8Reshaping Gradients
GRASP reshapes gradients to ensure actions receive clean signals while avoiding brittle 'state-input' gradients through high-dimensional vision models.
9Application Areas
GRASP can be used in robotics, autonomous vehicles, and other fields requiring long-horizon planning.
10Future Potential
GRASP has the potential to expand the use of learned world models by making long-horizon planning more practical and robust.
11Conclusion
GRASP provides an effective solution to the challenges of long-horizon planning and holds significant potential for future applications.