UAP-BEV: Uncertainty-Aware Planning in Bird's Eye View representations from Surround Monocular Images
Vikrant Dewangan [1], Basant Sharma [2], Tushar Choudhary [1], Sarthak Sharma [1], Aakash Aanegola [1],
Vikrant Dewangan [1], Basant Sharma [2], Tushar Choudhary [1], Sarthak Sharma [1], Aakash Aanegola [1],
(1- Robotics Research Center, IIIT Hyderabad, 2- University of Tartu, Estonia)
International Conference on Automation, Science and Engineering (CASE) 2023
Given is scene simulated in CARLA. (a): The ego vehicle (blue) is driving and has to overtake the static vehicle in front to reach destination using a BEV perception model ST-P3. (b) ST-P3 Gaussian: Using a Gaussian approximation of underlying uncertainty proves to be conservative and ego-vehicle fails to move ahead. (c) ST-P3 IL Traditional Imitation-Learning (IL) based approaches cannot account for the error and result in collision. Our uncertainty aware planner (d): UAP-BEV is able to overtake the leading vehicle while countering the error in perception, maintaining a safe distance compared to others.
The mean and variance of the Error Distribution with time for CARLA and NuScenes. increase with time, indicating that predictions further into the future are less reliable. This motivates us to account for uncertainty in perception while computing optimal plans.
Our approach uses a Spatio-Temporal Network ST-P3 to obtain a set of BEV predictions into the future which is then converted into an occupancy map prediction. We used the ground-truth information to learn the uncertainty in the occupancy map prediction. During inference time, we query the closest occupied cell to the ego-vehicle and then perturb it with samples drawn from the learned uncertainty. We then use the noisy samples of distance queries and use Reproducing Kernel Hilbert Spaces (RKHS) of Probability Density Functions (PDF) of Collision Violation Function, to optimise our uncertainty-aware trajectory with the Maximum Mean discrepancy (MMD) measure as the surrogate cost for collision avoidance. We adopt a sampling-based approach and augment a projection operator into the optimization pipeline for constraint satisfaction.
Given static vehicles placed along the route, ego-vehicle to perform multiple overtaking to reach the destination.
Ego-vehicle experiences cut-in from vehicle in adjacent lane. Ego-vehicle has to slow down to prevent collision.
Ego-vehicle performing a turn at an uncontrolled intersection and avoids collision with another vehicle.
A complete run of the simulation
Results
CARLA
Given is a cutin scenario generated through the ST-P3 pipeline. (a) represents the error in perception that occurs for the leading vehicle during the cut-in. In Fig.(b) it can be seen there is a difference between the ground truth (d_{GT}) and the predicted (d_{pred}) distance to the closest occupied cell at different time instants. As a result, the ST-P3 trajectory leads to a collision, denoted by the d_{GT} going below r_{safe}. In contrast, our approach successfully navigates the cutin. This result is visualized through scene images in Fig. (c)
Our UAP-BEV shows almost two to three times reduction in a collision over LSS and ST-P3-IL. Similar improvements can be found in the smoothness metric as well. We recall that NuScenes dataset only allows us to perform open-loop simulation (executed trajectory of ego-vehicle is pre-decided and fixed). Thus the improvement here is less drastic compared to CARLA. We also do not report the route-competition metric as it is irrelevant here.
We perform Convergence Analysis on our CEM method.
The cost profile of the mean trajectory is plotted and can be seen in . This signifies that the gradient update to the behavioral inputs distribution is such that the trajectories coming from it correspond to the lower cost regions.
Longitudinal Barrier Ablation
We demonstrate the effectiveness of the longitudinal barrier constraints. This constraint ensures a minimum separation distance from the leading vehicle in the scene. Note that we only utilize this constraint during the Inlane driving scenarios. The trajectory of the leading vehicle was obtained by computing the approximate centres of the BEV predictions.