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 Filos McAllister Rhinehart Levine Gal equal
Angelos Filos∗α Panos Tigkas∗α Rowan McAllisterβ Nick Rhinehartβ Sergey Levineβ Yarin Galα
∗ equal contribution αUniversity of Oxford βUniversity of California, Berkeley
Learning from Demonstrations
Training Time:
demo
xi yi
N i 1
Evaluation Time: navigate safely to goals
Sensitivity to Distribution Shifts (OOD)
in-distribution scenes
Decision-making under distribution shift
Learning from Demonstrations
Training Time:
demo
xi yi
N i 1
Evaluation Time: navigate safely to goals
Sensitivity to Distribution Shifts (OOD)
in-distribution scenes
Decision-making under distribution shift
Learning from Demonstrations
Training Time: Ddemo = {(xi, yi)}N
i=1
Evaluation Time: navigate safely to goals
Sensitivity to Distribution Shifts (OOD)
in-distribution scenes
Decision-making under distribution shift
Learning from Demonstrations
Training Time: Ddemo = {(xi, yi)}N
i=1
Evaluation Time: navigate safely to goals
Sensitivity to Distribution Shifts (OOD)
in-distribution scenes
Decision-making under distribution shift
Learning from Demonstrations
Training Time: Ddemo = {(xi, yi)}N
i=1
Evaluation Time: navigate safely to goals
Sensitivity to Distribution Shifts (OOD)
in-distribution scenes
Decision-making under distribution shift
Learning from Demonstrations
Training Time: Ddemo = {(xi, yi)}N
i=1
Evaluation Time: navigate safely to goals
Sensitivity to Distribution Shifts (OOD)
in-distribution scenes
Decision-making under distribution shift
Learning from Demonstrations
Training Time: Ddemo = {(xi, yi)}N
i=1
Evaluation Time: navigate safely to goals
Sensitivity to Distribution Shifts (OOD)
in-distribution scenes
Decision-making under distribution shift
Learning from Demonstrations
Training Time: Ddemo = {(xi, yi)}N
i=1
Evaluation Time: navigate safely to goals
Sensitivity to Distribution Shifts (OOD)
in-distribution scenes
Decision-making under distribution shift
Learning from Demonstrations
Training Time: Ddemo = {(xi, yi)}N
i=1
Evaluation Time: navigate safely to goals
Sensitivity to Distribution Shifts (OOD)
in-distribution scenes
Decision-making under distribution shift
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.
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.
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
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
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
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
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
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
CARLA @ Roundabout OOD driving scene
models, qk trajectories, yi y1 y2 y3 q1 0.6 0.1 0.3 q2 0.3 0.4 0.3 q3 0.2 0.2 0.6 mink 0.2 0.1 0.3 y3
1 K ∑k 1.1 3 0.7 3 1.2 3
y3 arg maxi y3 y2 y1
Robust Imitative Planning “plan” ← “aggregate” ← “evaluate”
Online Planning Under Epistemic Uncertainty
CARLA @ Roundabout OOD driving scene
models, qk trajectories, yi y1 y2 y3 q1 0.6 0.1 0.3 q2 0.3 0.4 0.3 q3 0.2 0.2 0.6 mink 0.2 0.1 0.3 y3
1 K ∑k 1.1 3 0.7 3 1.2 3
y3 arg maxi y3 y2 y1
Robust Imitative Planning “plan” ← “aggregate” ← “evaluate”
Online Planning Under Epistemic Uncertainty
CARLA @ Roundabout OOD driving scene
models, qk trajectories, yi y1 y2 y3 q1 0.6 0.1 0.3 q2 0.3 0.4 0.3 q3 0.2 0.2 0.6 mink 0.2 0.1 0.3 y3
1 K ∑k 1.1 3 0.7 3 1.2 3
y3 arg maxi y3 y2 y1
Robust Imitative Planning “plan” ← “aggregate” ← “evaluate”
Online Planning Under Epistemic Uncertainty
CARLA @ Roundabout OOD driving scene
models, qk trajectories, yi y1 y2 y3 q1 0.6 0.1 0.3 q2 0.3 0.4 0.3 q3 0.2 0.2 0.6 mink 0.2 0.1 0.3 y3
1 K ∑k 1.1 3 0.7 3 1.2 3
y3 arg maxi y3 y2 y1
Robust Imitative Planning “plan” ← “aggregate” ← “evaluate”
Online Planning Under Epistemic Uncertainty
CARLA @ Roundabout OOD driving scene
models, qk trajectories, yi y1 y2 y3 q1 0.6 0.1 0.3 q2 0.3 0.4 0.3 q3 0.2 0.2 0.6 mink 0.2 0.1 0.3 y3
1 K ∑k 1.1 3 0.7 3 1.2 3
y3 arg maxi y3 y2 y1
Robust Imitative Planning “plan” ← “aggregate” ← “evaluate”
Online Planning Under Epistemic Uncertainty
CARLA @ Roundabout OOD driving scene
models, qk trajectories, yi y1 y2 y3 q1 0.6 0.1 0.3 q2 0.3 0.4 0.3 q3 0.2 0.2 0.6 mink 0.2 0.1 0.3 y3
1 K ∑k 1.1 3 0.7 3 1.2 3
y3 arg maxi y3 y2 y1
Robust Imitative Planning “plan” ← “aggregate” ← “evaluate”
Online Planning Under Epistemic Uncertainty
CARLA @ Roundabout OOD driving scene
models, qk trajectories, yi y1 y2 y3 q1 0.6 0.1 0.3 q2 0.3 0.4 0.3 q3 0.2 0.2 0.6 mink 0.2 0.1 0.3 y3
1 K ∑k 1.1 3 0.7 3 1.2 3
y3 arg maxi y3 y2 y1
Robust Imitative Planning “plan” ← “aggregate” ← “evaluate”
Online Planning Under Epistemic Uncertainty
CARLA @ Roundabout OOD driving scene
models, qk trajectories, yi y1 y2 y3 q1 0.6 0.1 0.3 q2 0.3 0.4 0.3 q3 0.2 0.2 0.6 mink 0.2 0.1 0.3 y3
1 K ∑k 1.1 3 0.7 3 1.2 3
y3 arg maxi y3 y2 y1
Robust Imitative Planning “plan” ← “aggregate” ← “evaluate”
Online Planning Under Epistemic Uncertainty
CARLA @ Roundabout OOD driving scene
models, qk trajectories, yi y1 y2 y3 q1 0.6 0.1 0.3 q2 0.3 0.4 0.3 q3 0.2 0.2 0.6 mink 0.2 0.1 0.3 y3
1 K ∑k 1.1 3 0.7 3 1.2 3
y3 arg maxi y3 y2 y1
Robust Imitative Planning “plan” ← “aggregate” ← “evaluate”
Online Planning Under Epistemic Uncertainty
CARLA @ Roundabout OOD driving scene
models, qk trajectories, yi y1 y2 y3 q1 0.6 0.1 0.3 q2 0.3 0.4 0.3 q3 0.2 0.2 0.6 mink 0.2 0.1 0.3 y3
1 K ∑k 1.1 3 0.7 3 1.2 3
y3 arg maxi y3 y2 y1
Robust Imitative Planning “plan” ← “aggregate” ← “evaluate”
Online Planning Under Epistemic Uncertainty yRIP-⊕ ≜ arg max
y
⊕
θ∈supp(p(θ|D))
log q(y|x; θ)
CARLA @ Roundabout OOD driving scene
models, qk trajectories, yi y1 y2 y3 q1 0.6 0.1 0.3 q2 0.3 0.4 0.3 q3 0.2 0.2 0.6 mink 0.2 0.1 0.3 y3
1 K ∑k 1.1 3 0.7 3 1.2 3
y3 arg maxi y3 y2 y1
Robust Imitative Planning “plan” ← “aggregate” ← “evaluate”
Online Planning Under Epistemic Uncertainty yRIP-WCM ≜ arg max
y
min
θ∈supp(p(θ|D))log q(y|x; θ)
CARLA @ Roundabout OOD driving scene
models, qk trajectories, yi y1 y2 y3 q1 0.6 0.1 0.3 q2 0.3 0.4 0.3 q3 0.2 0.2 0.6 mink 0.2 0.1 0.3 y3
1 K ∑k 1.1 3 0.7 3 1.2 3
y3 arg maxi y3 y2 y1
Robust Imitative Planning “plan” ← “aggregate” ← “evaluate”
Online Planning Under Epistemic Uncertainty sRIP-MA ≜ arg max
y
∫
p(θ|D)log q(y|x; θ)dθ
Algorithm 1: Robust Imitaive Planning (RIP) // ``train'': {θk}K
k=1 1 learn from demonstrations with MLE 2 for step in environment do
// ``init'': y
3
sample random plan
4
while note converged do
5
// ``evaluate'': {q(y|x, G; θk)}K
1 6
score plan under imitative model(s)
7
// ``aggregate'': ⊕, mink,
1 K ∑k 8
consolidate evaluations of ensemble
9
// ``plan'': y ← y + η ∂U
∂y 10
improve plan with online SGD // ``act'': y∗
11
submit plan to environment
K ∑k
∂U ∂y
Algorithm 2: Robust Imitaive Planning (RIP) // ``train'': {θk}K
k=1 1 learn from demonstrations with MLE 2 for step in environment do
// ``init'': y
3
sample random plan
4
while note converged do
5
// ``evaluate'': {q(y|x, G; θk)}K
1 6
score plan under imitative model(s)
7
// ``aggregate'': ⊕, mink,
1 K ∑k 8
consolidate evaluations of ensemble
9
// ``plan'': y ← y + η ∂U
∂y 10
improve plan with online SGD // ``act'': y∗
11
submit plan to environment
K ∑k
∂U ∂y
Algorithm 3: Robust Imitaive Planning (RIP) // ``train'': {θk}K
k=1 1 learn from demonstrations with MLE 2 for step in environment do
// ``init'': y
3
sample random plan
4
while note converged do
5
// ``evaluate'': {q(y|x, G; θk)}K
1 6
score plan under imitative model(s)
7
// ``aggregate'': ⊕, mink,
1 K ∑k 8
consolidate evaluations of ensemble
9
// ``plan'': y ← y + η ∂U
∂y 10
improve plan with online SGD // ``act'': y∗
11
submit plan to environment
K ∑k
∂U ∂y
Algorithm 4: Robust Imitaive Planning (RIP) // ``train'': {θk}K
k=1 1 learn from demonstrations with MLE 2 for step in environment do
// ``init'': y
3
sample random plan
4
while note converged do
5
// ``evaluate'': {q(y|x, G; θk)}K
1 6
score plan under imitative model(s)
7
// ``aggregate'': ⊕, mink,
1 K ∑k 8
consolidate evaluations of ensemble
9
// ``plan'': y ← y + η ∂U
∂y 10
improve plan with online SGD // ``act'': y∗
11
submit plan to environment
K ∑k
∂U ∂y
Algorithm 5: Robust Imitaive Planning (RIP) // ``train'': {θk}K
k=1 1 learn from demonstrations with MLE 2 for step in environment do
// ``init'': y
3
sample random plan
4
while note converged do
5
// ``evaluate'': {q(y|x, G; θk)}K
1 6
score plan under imitative model(s)
7
// ``aggregate'': ⊕, mink,
1 K ∑k 8
consolidate evaluations of ensemble
9
// ``plan'': y ← y + η ∂U
∂y 10
improve plan with online SGD // ``act'': y∗
11
submit plan to environment
K ∑k
∂U ∂y
Algorithm 6: Robust Imitaive Planning (RIP) // ``train'': {θk}K
k=1 1 learn from demonstrations with MLE 2 for step in environment do
// ``init'': y
3
sample random plan
4
while note converged do
5
// ``evaluate'': {q(y|x, G; θk)}K
1 6
score plan under imitative model(s)
7
// ``aggregate'': ⊕, mink,
1 K ∑k 8
consolidate evaluations of ensemble
9
// ``plan'': y ← y + η ∂U
∂y 10
improve plan with online SGD // ``act'': y∗
11
submit plan to environment
K ∑k
∂U ∂y
Algorithm 7: Robust Imitaive Planning (RIP) // ``train'': {θk}K
k=1 1 learn from demonstrations with MLE 2 for step in environment do
// ``init'': y
3
sample random plan
4
while note converged do
5
// ``evaluate'': {q(y|x, G; θk)}K
1 6
score plan under imitative model(s)
7
// ``aggregate'': ⊕, mink,
1 K ∑k 8
consolidate evaluations of ensemble
9
// ``plan'': y ← y + η ∂U
∂y 10
improve plan with online SGD // ``act'': y∗
11
submit plan to environment
K ∑k
∂U ∂y
ICRA 2020 nuScenes prediction challenge Boston Singapore minADE1 ↓ minADE5 ↓ minFDE1 ↓ minADE1 ↓ minADE5 ↓ minFDE1 ↓ Methods (2073 scenes, 50 samples, open-loop planning) (1189 scenes, 50 samples, open-loop planning) MTP♦⋆ (Cui et al., 2019) 4.13 3.24 9.23 4.13 3.24 9.23 MultiPath♦⋆ (Chai et al., 2019) 3.89 3.34 9.19 3.89 3.34 9.19 CoverNet♦⋆ (Phan-Minh et al., 2019) 3.87 2.41 9.26 3.87 2.41 9.26 DIM♣ (Rhinehart et al., 2020) 3.64±0.05 2.48±0.02 8.22±0.13 3.82±0.04 2.95±0.01 8.91±0.08 RIP-BCM♣ (baseline) 3.53±0.04 2.37±0.01 7.92±0.09 3.57±0.02 2.70±0.01 8.39±0.03 RIP-MA♣ (ours) 3.39±0.03 2.33±0.01 7.62±0.07 3.48±0.01 2.69±0.02 8.19±0.02 RIP-WCM♣ (ours) 3.29±0.03 2.28±0.00 7.45±0.05 3.43±0.01 2.66±0.01 8.09±0.04
CARNOVEL benchmark of novel driving scenarios (based on CARLA) AbnormalTurns BusyTown Success ↑ Infra/km ↓ Success ↑ Infra/km ↓ Methods (7 × 10 scenes, %) (×1e − 3) (11 × 10 scenes, %) (×1e − 3) CIL♣⋆ (Codevilla et al., 2018) 65.71±7.37 7.04±5.07 5.45±6.35 11.49±3.66 LbC♣⋆ (Chen et al., 2019) 0.00±0.00 5.81±0.58 20.00±13.48 3.96±0.15 DIM♣ (Rhinehart et al., 2020) 74.28±11.26 5.56±4.06 47.13±14.54 8.47±5.22 RIP-BCM♣ (baseline) 68.57±9.03 7.93±3.73 50.90±20.64 3.74±5.52 RIP-MA♣ (ours) 84.28±14.20 7.86±5.70 64.54±23.25 5.86±3.99 RIP-WCM♣ (ours) 87.14±14.20 4.91±3.60 62.72±5.16 3.17±2.04
ICRA 2020 nuScenes prediction challenge Boston Singapore minADE1 ↓ minADE5 ↓ minFDE1 ↓ minADE1 ↓ minADE5 ↓ minFDE1 ↓ Methods (2073 scenes, 50 samples, open-loop planning) (1189 scenes, 50 samples, open-loop planning) MTP♦⋆ (Cui et al., 2019) 4.13 3.24 9.23 4.13 3.24 9.23 MultiPath♦⋆ (Chai et al., 2019) 3.89 3.34 9.19 3.89 3.34 9.19 CoverNet♦⋆ (Phan-Minh et al., 2019) 3.87 2.41 9.26 3.87 2.41 9.26 DIM♣ (Rhinehart et al., 2020) 3.64±0.05 2.48±0.02 8.22±0.13 3.82±0.04 2.95±0.01 8.91±0.08 RIP-BCM♣ (baseline) 3.53±0.04 2.37±0.01 7.92±0.09 3.57±0.02 2.70±0.01 8.39±0.03 RIP-MA♣ (ours) 3.39±0.03 2.33±0.01 7.62±0.07 3.48±0.01 2.69±0.02 8.19±0.02 RIP-WCM♣ (ours) 3.29±0.03 2.28±0.00 7.45±0.05 3.43±0.01 2.66±0.01 8.09±0.04
CARNOVEL benchmark of novel driving scenarios (based on CARLA) AbnormalTurns BusyTown Success ↑ Infra/km ↓ Success ↑ Infra/km ↓ Methods (7 × 10 scenes, %) (×1e − 3) (11 × 10 scenes, %) (×1e − 3) CIL♣⋆ (Codevilla et al., 2018) 65.71±7.37 7.04±5.07 5.45±6.35 11.49±3.66 LbC♣⋆ (Chen et al., 2019) 0.00±0.00 5.81±0.58 20.00±13.48 3.96±0.15 DIM♣ (Rhinehart et al., 2020) 74.28±11.26 5.56±4.06 47.13±14.54 8.47±5.22 RIP-BCM♣ (baseline) 68.57±9.03 7.93±3.73 50.90±20.64 3.74±5.52 RIP-MA♣ (ours) 84.28±14.20 7.86±5.70 64.54±23.25 5.86±3.99 RIP-WCM♣ (ours) 87.14±14.20 4.91±3.60 62.72±5.16 3.17±2.04
ICRA 2020 nuScenes prediction challenge Boston Singapore minADE1 ↓ minADE5 ↓ minFDE1 ↓ minADE1 ↓ minADE5 ↓ minFDE1 ↓ Methods (2073 scenes, 50 samples, open-loop planning) (1189 scenes, 50 samples, open-loop planning) MTP♦⋆ (Cui et al., 2019) 4.13 3.24 9.23 4.13 3.24 9.23 MultiPath♦⋆ (Chai et al., 2019) 3.89 3.34 9.19 3.89 3.34 9.19 CoverNet♦⋆ (Phan-Minh et al., 2019) 3.87 2.41 9.26 3.87 2.41 9.26 DIM♣ (Rhinehart et al., 2020) 3.64±0.05 2.48±0.02 8.22±0.13 3.82±0.04 2.95±0.01 8.91±0.08 RIP-BCM♣ (baseline) 3.53±0.04 2.37±0.01 7.92±0.09 3.57±0.02 2.70±0.01 8.39±0.03 RIP-MA♣ (ours) 3.39±0.03 2.33±0.01 7.62±0.07 3.48±0.01 2.69±0.02 8.19±0.02 RIP-WCM♣ (ours) 3.29±0.03 2.28±0.00 7.45±0.05 3.43±0.01 2.66±0.01 8.09±0.04
CARNOVEL benchmark of novel driving scenarios (based on CARLA) AbnormalTurns BusyTown Success ↑ Infra/km ↓ Success ↑ Infra/km ↓ Methods (7 × 10 scenes, %) (×1e − 3) (11 × 10 scenes, %) (×1e − 3) CIL♣⋆ (Codevilla et al., 2018) 65.71±7.37 7.04±5.07 5.45±6.35 11.49±3.66 LbC♣⋆ (Chen et al., 2019) 0.00±0.00 5.81±0.58 20.00±13.48 3.96±0.15 DIM♣ (Rhinehart et al., 2020) 74.28±11.26 5.56±4.06 47.13±14.54 8.47±5.22 RIP-BCM♣ (baseline) 68.57±9.03 7.93±3.73 50.90±20.64 3.74±5.52 RIP-MA♣ (ours) 84.28±14.20 7.86±5.70 64.54±23.25 5.86±3.99 RIP-WCM♣ (ours) 87.14±14.20 4.91±3.60 62.72±5.16 3.17±2.04
Algorithm 7: Adaptive RIP (AdaRIP) // ``train'': {θk}K
k=1 1 learn from demonstrations with MLE 2 for step in environment do
// ``init'': y sample random plan while note converged do // ``evaluate'': {q(y|x, G; θk)}K
1 3
score plan under imitative model(s) // ``aggregate'': ⊕, mink,
1 K ∑k 4
consolidate evaluations of ensemble // ``plan'': y ← y + η ∂U
∂y 5
improve plan with online SGD // ``adapt'': u(y∗) > τ
6
if epistemically uncertain then
7
query expert
8
update model to reduce uncertainty // ``act'': y∗
9
submit plan to environment
Algorithm 8: Adaptive RIP (AdaRIP) // ``train'': {θk}K
k=1 1 learn from demonstrations with MLE 2 for step in environment do
// ``init'': y sample random plan while note converged do // ``evaluate'': {q(y|x, G; θk)}K
1 3
score plan under imitative model(s) // ``aggregate'': ⊕, mink,
1 K ∑k 4
consolidate evaluations of ensemble // ``plan'': y ← y + η ∂U
∂y 5
improve plan with online SGD // ``adapt'': u(y∗) > τ
6
if epistemically uncertain then
7
query expert
8
update model to reduce uncertainty // ``act'': y∗
9
submit plan to environment
Algorithm 9: Adaptive RIP (AdaRIP) // ``train'': {θk}K
k=1 1 learn from demonstrations with MLE 2 for step in environment do
// ``init'': y sample random plan while note converged do // ``evaluate'': {q(y|x, G; θk)}K
1 3
score plan under imitative model(s) // ``aggregate'': ⊕, mink,
1 K ∑k 4
consolidate evaluations of ensemble // ``plan'': y ← y + η ∂U
∂y 5
improve plan with online SGD // ``adapt'': u(y∗) > τ
6
if epistemically uncertain then
7
query expert
8
update model to reduce uncertainty // ``act'': y∗
9
submit plan to environment
(Normalized) Uncertainty RIP AdaRIP
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
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
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
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
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
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
Angelos Filos∗α Panos Tigkas∗α Rowan McAllisterβ Nick Rhinehartβ Sergey Levineβ Yarin Galα
∗ equal contribution αUniversity of Oxford βUniversity of California, Berkeley
Codevilla, Felipe et al. (2018). “End-to-end driving via conditional imitation learning”. International Conference on Robotics and Automation (ICRA). IEEE, pp. 1–9. Chai, Yuning et al. (2019). “Multipath: Multiple probabilistic anchor trajectory hypotheses for behavior prediction”. arXiv preprint arXiv:1910.05449. Chen, Dian et al. (2019). “Learning by cheating”. arXiv preprint arXiv:1912.12294. Cui, Henggang et al. (2019). “Multimodal trajectory predictions for autonomous driving using deep convolutional networks”. 2019 International Conference on Robotics and Automation (ICRA). IEEE, pp. 2090–2096. Phan-Minh, Tung et al. (2019). “CoverNet: Multimodal Behavior Prediction using Trajectory Sets”. arXiv preprint arXiv:1911.10298. Rhinehart, Nicholas, Rowan McAllister, and Sergey Levine (2020). “Deep Imitative Models for Flexible Inference, Planning, and Control”. International Conference on Learning Representations (ICLR).