Khái niệm cốt lõi
A "generate-and-test" approach to risk-bounded planning for autonomous mobile agents with learned, non-linear, stochastic dynamics. The method uses a variational autoencoder to learn an approximate linear latent dynamics model and performs trajectory optimization in the latent space, while a validator assesses the risk of the candidate trajectory and computes additional safety constraints.
Tóm tắt
The paper introduces Latent Space Planning for Stochastic Systems (LaPlaSS), an algorithm for risk-bounded planning for non-linear stochastic agents with learned dynamics. The approach employs a "generate-and-test" strategy:
- The trajectory planner generates a candidate control trajectory using an approximate linear dynamics model in the latent space learned via a variational autoencoder (VAE).
- The trajectory validator samples trajectories around the candidate using the accurate, stochastic VAE dynamics model to compute a probabilistic flow tube. It then assesses the risk of collision with obstacles and computes a safety constraint if the risk bound is violated.
- The safety constraint is passed back to the trajectory planner, which refines the candidate trajectory in the next iteration.
- The process iterates until a safe trajectory is found or the problem is deemed infeasible.
The key components are:
- Learning an accurate stochastic dynamics model using a VAE to capture uncertainty.
- Learning an approximate linear dynamics model in the latent space of the VAE for efficient trajectory optimization.
- Encoding the planning problem into the latent space to leverage convex optimization techniques.
- The "generate-and-test" approach that iterates between planning and validation to ensure risk-bounded trajectories.
The authors demonstrate that their algorithm can generate safe trajectories an order of magnitude faster than the state-of-the-art approach for agents with known dynamics, and show successful planning for a real-world autonomous drone with learned dynamics.
Thống kê
The authors evaluate their approach on a simulated non-linear stochastic agent with known dynamics, as well as a real-world autonomous drone dataset.
For the simulated agent:
The average time to generate a safe trajectory was 11 seconds, compared to over 120 seconds for the state-of-the-art approach.
The average total risk of the solution was 0.032, compared to 0.023 for the state-of-the-art.
The average objective value (sum of squared velocity and steering angle) was 85.28, compared to 74.16 for the state-of-the-art.
For the real-world drone dataset:
The authors successfully generated a safe trajectory plan for the quadrotor without prior knowledge of the dynamics.
The final trajectory respected the control limits, avoided ellipsoidal obstacles, and took the agent from the initial state to the goal.