<ul><li>Continuous time systems are typically modeled using discrete time dynamics, requiring a small simulation step for accuracy.</li><li>Proposed approach involves using temporally-extended actions to control continuous decision timescale, allowing for deep horizon search with shallow planner depth.</li><li>This method speeds up trajectory simulation, reduces errors in model-based reinforcement learning, and improves training time for models.</li><li>Experimental evaluation demonstrates that the approach results in faster planning, better solutions, and solves problems not addressed in the standard formulation.</li></ul>

Improving planning and MBRL with temporally-extended actions

Discover more