Can Autonomous Vehicles Identify, Recover From, Rowan - PowerPoint PPT Presentation

Can Autonomous Vehicles Identify, Recover From, Rowan github.com/OATML/carsuite Yarin Sergey and Adapt to Distribution Shifts? Nick Angelos Panos ICML 2020 Tigkas ∗ α Filos ∗ α McAllister β Rhinehart β Levine β Gal α ∗ equal contribution α University of Oxford β University of California, Berkeley

x i y i navigate safely to goals Problem Setting Evaluation Time : in-distribution scenes out-of-distribution (OOD) scenes • inexhaustible space of scenes • failure to generalise in OOD Sensitivity to Distribution Shifts (OOD) 1 Learning from Demonstrations i N demo Training Time : • plethora of expert demonstrations • no explicit reward function Decision-making under distribution shift

x i y i navigate safely to goals Problem Setting Evaluation Time : in-distribution scenes out-of-distribution (OOD) scenes • inexhaustible space of scenes • failure to generalise in OOD Sensitivity to Distribution Shifts (OOD) 1 Learning from Demonstrations i N demo Training Time : • plethora of expert demonstrations • no explicit reward function Decision-making under distribution shift

Problem Setting Learning from Demonstrations • no explicit reward function • plethora of expert demonstrations Training Time : Evaluation Time : Sensitivity to Distribution Shifts (OOD) • failure to generalise in OOD • inexhaustible space of scenes out-of-distribution (OOD) scenes in-distribution scenes Decision-making under distribution shift D demo = { ( x i , y i ) } N i = 1 navigate safely to goals

Problem Setting Learning from Demonstrations • no explicit reward function • plethora of expert demonstrations Training Time : Evaluation Time : Sensitivity to Distribution Shifts (OOD) • failure to generalise in OOD • inexhaustible space of scenes out-of-distribution (OOD) scenes in-distribution scenes Decision-making under distribution shift D demo = { ( x i , y i ) } N i = 1 navigate safely to goals

Problem Setting Learning from Demonstrations • no explicit reward function • plethora of expert demonstrations Training Time : Evaluation Time : Sensitivity to Distribution Shifts (OOD) • failure to generalise in OOD • inexhaustible space of scenes out-of-distribution (OOD) scenes in-distribution scenes Decision-making under distribution shift D demo = { ( x i , y i ) } N i = 1 navigate safely to goals

Problem Setting Learning from Demonstrations • no explicit reward function • plethora of expert demonstrations Training Time : Evaluation Time : Sensitivity to Distribution Shifts (OOD) • failure to generalise in OOD • inexhaustible space of scenes out-of-distribution (OOD) scenes in-distribution scenes Decision-making under distribution shift D demo = { ( x i , y i ) } N i = 1 navigate safely to goals

Problem Setting Learning from Demonstrations • no explicit reward function • plethora of expert demonstrations Training Time : Evaluation Time : Sensitivity to Distribution Shifts (OOD) • failure to generalise in OOD • inexhaustible space of scenes out-of-distribution (OOD) scenes in-distribution scenes Decision-making under distribution shift D demo = { ( x i , y i ) } N i = 1 navigate safely to goals

Problem Setting Learning from Demonstrations • no explicit reward function • plethora of expert demonstrations Training Time : Evaluation Time : Sensitivity to Distribution Shifts (OOD) • failure to generalise in OOD • inexhaustible space of scenes out-of-distribution (OOD) scenes in-distribution scenes Decision-making under distribution shift D demo = { ( x i , y i ) } N i = 1 navigate safely to goals

Problem Setting Learning from Demonstrations • no explicit reward function • plethora of expert demonstrations Training Time : Evaluation Time : Sensitivity to Distribution Shifts (OOD) • failure to generalise in OOD • inexhaustible space of scenes out-of-distribution (OOD) scenes in-distribution scenes Decision-making under distribution shift D demo = { ( x i , y i ) } N i = 1 navigate safely to goals

Main Result CARLA @ Roundabout nuScenes @ Boston Nicholas Rhinehart et al. (2020). “Deep Imitative Models for Flexible Inference, Planning, and Control”. International Conference on Learning Representations (ICLR) . Felipe Codevilla et al. (2018). “End-to-end driving via conditional imitation learning”. International Conference on Robotics and Automation (ICRA) . IEEE, pp. 1–9.

Main Result Uncertainty-Aware Online Planning in OOD Nicholas Rhinehart et al. (2020). “Deep Imitative Models for Flexible Inference, Planning, and Control”. International Conference on Learning Representations (ICLR) . Felipe Codevilla et al. (2018). “End-to-end driving via conditional imitation learning”. International Conference on Robotics and Automation (ICRA) . IEEE, pp. 1–9.

Contributions Identify novel (OOD) scenes in-distribution scenes Epistemic (Model) Uncertainty Recover From Robust Imitative Planning (RIP) Adapt To Adaptive RIP (AdaRIP) Autonomous car novel-scene benchmark CARNOVEL github.com/OATML/carsuite AbnormnalTurns BusyTown Hills Roundabouts

Contributions Identify novel (OOD) scenes in-distribution scenes Epistemic (Model) Uncertainty Recover From Robust Imitative Planning (RIP) Adapt To Adaptive RIP (AdaRIP) Autonomous car novel-scene benchmark CARNOVEL github.com/OATML/carsuite AbnormnalTurns BusyTown Hills Roundabouts

Contributions Identify novel (OOD) scenes in-distribution scenes Epistemic (Model) Uncertainty Recover From Robust Imitative Planning (RIP) Adapt To Adaptive RIP (AdaRIP) Autonomous car novel-scene benchmark CARNOVEL github.com/OATML/carsuite AbnormnalTurns BusyTown Hills Roundabouts

Contributions Identify novel (OOD) scenes in-distribution scenes Epistemic (Model) Uncertainty Recover From Robust Imitative Planning (RIP) Adapt To Adaptive RIP (AdaRIP) Autonomous car novel-scene benchmark CARNOVEL github.com/OATML/carsuite AbnormnalTurns BusyTown Hills Roundabouts

Contributions Identify novel (OOD) scenes in-distribution scenes Epistemic (Model) Uncertainty Recover From Robust Imitative Planning (RIP) Adapt To Adaptive RIP (AdaRIP) Autonomous car novel-scene benchmark CARNOVEL github.com/OATML/carsuite AbnormnalTurns BusyTown Hills Roundabouts

Contributions Identify novel (OOD) scenes in-distribution scenes Epistemic (Model) Uncertainty Recover From Robust Imitative Planning (RIP) Adapt To Adaptive RIP (AdaRIP) Autonomous car novel-scene benchmark CARNOVEL github.com/OATML/carsuite AbnormnalTurns BusyTown Hills Roundabouts

Recover From OOD: Robust Imitative Planning (RIP) 3 CARLA @ Roundabout 0.2 0.1 0.3 y 3 1 3 0.2 3 y 3 y 3 y 2 y 1 Robust Imitative Planning 0.6 Online Planning Under Epistemic Uncertainty 0.2 models, q k y 2 y 3 trajectories, y i q 1 q 3 0.6 OOD driving scene y 1 0.1 0.3 q 2 0.3 0.4 0.3 arg max i min k 1 . 1 0 . 7 1 . 2 K ∑ k “plan” ← “aggregate” ← “evaluate”

Recover From OOD: Robust Imitative Planning (RIP) 3 CARLA @ Roundabout 0.2 0.1 0.3 y 3 1 3 0.2 3 y 3 y 3 y 2 y 1 Robust Imitative Planning 0.6 Online Planning Under Epistemic Uncertainty 0.2 models, q k y 2 y 3 trajectories, y i q 1 q 3 0.6 OOD driving scene y 1 0.1 0.3 q 2 0.3 0.4 0.3 arg max i min k 1 . 1 0 . 7 1 . 2 K ∑ k “plan” ← “aggregate” ← “evaluate”

Recover From OOD: Robust Imitative Planning (RIP) 3 CARLA @ Roundabout 0.2 0.1 0.3 y 3 1 3 0.2 3 y 3 y 3 y 2 y 1 Robust Imitative Planning 0.6 Online Planning Under Epistemic Uncertainty 0.2 models, q k y 2 y 3 trajectories, y i q 1 q 3 0.6 OOD driving scene y 1 0.1 0.3 q 2 0.3 0.4 0.3 arg max i min k 1 . 1 0 . 7 1 . 2 K ∑ k “plan” ← “aggregate” ← “evaluate”

Recover From OOD: Robust Imitative Planning (RIP) 3 CARLA @ Roundabout 0.2 0.1 0.3 y 3 1 3 0.2 3 y 3 y 3 y 2 y 1 Robust Imitative Planning 0.6 Online Planning Under Epistemic Uncertainty 0.2 models, q k y 2 y 3 trajectories, y i q 1 q 3 0.6 OOD driving scene y 1 0.1 0.3 q 2 0.3 0.4 0.3 arg max i min k 1 . 1 0 . 7 1 . 2 K ∑ k “plan” ← “aggregate” ← “evaluate”

Recover From OOD: Robust Imitative Planning (RIP) 3 CARLA @ Roundabout 0.2 0.1 0.3 y 3 1 3 0.2 3 y 3 y 3 y 2 y 1 Robust Imitative Planning 0.6 Online Planning Under Epistemic Uncertainty 0.2 models, q k y 2 y 3 trajectories, y i q 1 q 3 0.6 OOD driving scene y 1 0.1 0.3 q 2 0.3 0.4 0.3 arg max i min k 1 . 1 0 . 7 1 . 2 K ∑ k “plan” ← “aggregate” ← “evaluate”

Recover From OOD: Robust Imitative Planning (RIP) 3 CARLA @ Roundabout 0.2 0.1 0.3 y 3 1 3 0.2 3 y 3 y 3 y 2 y 1 Robust Imitative Planning 0.6 Online Planning Under Epistemic Uncertainty 0.2 models, q k y 2 y 3 trajectories, y i q 1 q 3 0.6 OOD driving scene y 1 0.1 0.3 q 2 0.3 0.4 0.3 arg max i min k 1 . 1 0 . 7 1 . 2 K ∑ k “plan” ← “aggregate” ← “evaluate”

Recover From OOD: Robust Imitative Planning (RIP) 3 CARLA @ Roundabout 0.2 0.1 0.3 y 3 1 3 0.2 3 y 3 y 3 y 2 y 1 Robust Imitative Planning 0.6 Online Planning Under Epistemic Uncertainty 0.2 models, q k y 2 y 3 trajectories, y i q 1 q 3 0.6 OOD driving scene y 1 0.1 0.3 q 2 0.3 0.4 0.3 arg max i min k 1 . 1 0 . 7 1 . 2 K ∑ k “plan” ← “aggregate” ← “evaluate”