driving autonomy - intelligent vehicles
TRANSCRIPT
Driving Autonomy
Rob Weston
Machine Learning for Intelligent Transportation Systems
OXFORD ROBOTICS INSTITUTE
APPLIED ARTIFICIAL INTELLIGENCE LABA2I
…is mobile autonomy
Autonomous driving is fundamentally a systems challenge …
Why systems?
Challenges • An uncertain world • Imperfect sensors and maps • Ever changing environments (unseen locations,
weather, season)
Solutions must be… • auditable • safe • efficient (computation, energy)
Localisation Detection Tracking Planning Control
Camera, Lidar, Radar Etc.Driving commands
Systems allow us to move from this…
…is mobile autonomy
To this!
To this!
The Deep Learning Revolution
We can now accurately learn some pretty complex mappings…
… and execute these mappings very quickly indeed.
So, where does machine learning fit in?
Improving the tooling… • automatic data generation • versatile simulation, e.g.
• real -> sim • third-party behaviour
• …
Improving the platform… • sensor performance (and fusion) • situational awareness
• perception • prediction (behaviour) • planning (behaviour)
• …
The Deep Learning Revolution
We can now accurately learn some pretty complex mappings…
… and execute these mappings very quickly indeed.
Provided we have vast labelled datasets!
Integrating Machine Learning into the system creates some unique challenges
Learning from limited training labels
Adapting to new situations quickly
Managing uncertainty
Model validation
Model Size …
So where does machine learning fit in?
Leveraging the system for self-supervision…
End-to-End Tracking and Semantic Segmentation Using RNNs
Peter Ondruska, Julie Dequaire, Dominic Wang and Ingmar Posner
[Ondruska, Dequaire, Wang and Posner, “End-to-End Tracking and Semantic Segmentation Using RNNs”, RSS w/s DL in Robotics 2016]
Input Output
Deep Tracking
[Ondruska, Dequaire, Wang and Posner, “End-to-End Tracking and Semantic Segmentation Using RNNs”, RSS w/s DL in Robotics 2016]
Self-supervised learning of world dynamics (and then semantics)
Identifying Domain Knowledge What do we have? What do we know?
What do we have? What do we know?
Path Proposals for Urban AutonomyDan Barnes, Will Maddern and Ingmar Posner
[Barnes et al, “Find Your Own Way: Self-Supervised Segmentation of Path Proposals for Urban Autonomy”, ICRA, 2017]
[Barnes et al, “Find Your Own Way: Self-Supervised Segmentation of Path Proposals for Urban Autonomy”, ICRA, 2017]
[Barnes et al, “Find Your Own Way: Self-Supervised Segmentation of Path Proposals for Urban Autonomy”, ICRA, 2017]
Training is entirely self-supervised.[Barnes et al, “Find Your Own Way: Self-Supervised Segmentation of Path Proposals for Urban Autonomy”, ICRA, 2017]
Deployed with monocular camera only.[Barnes et al, “Find Your Own Way: Self-Supervised Segmentation of Path Proposals for Urban Autonomy”, ICRA, 2017]
Generating training dataPath Proposals
[Barnes et al, “Find Your Own Way: Self-Supervised Segmentation of Path Proposals for Urban Autonomy”, ICRA, 2017]
Generalises across appearance changes…
[Barnes et al, “Find Your Own Way: Self-Supervised Segmentation of Path Proposals for Urban Autonomy”, ICRA, 2017]
… and to new environments
[Barnes et al, “Find Your Own Way: Self-Supervised Segmentation of Path Proposals for Urban Autonomy”, ICRA, 2017]
Self-Supervised Distraction SuppressionDan Barnes, Will Maddern, Geoff Pascoe and Ingmar Posner
[Barnes et al, “Driven to Distraction: Self-Supervised Distractor Learning for Robust Monocular Visual Odometry in Urban Environments”, ICRA 2018]
Deep Distraction SuppressionLearning where to look…
[Barnes et al, “Driven to Distraction: Self-Supervised Distractor Learning for Robust Monocular Visual Odometry in Urban Environments”, ICRA 2018]
Deep Distraction SuppressionLearning where to look…
[Barnes et al, “Driven to Distraction: Self-Supervised Distractor Learning for Robust Monocular Visual Odometry in Urban Environments”, ICRA 2018]
[Barnes et al, “Driven to Distraction: Self-Supervised Distractor Learning for Robust Monocular Visual Odometry in Urban Environments”, ICRA 2018]
OverviewDeep Distraction Suppression (& Mono VO)
Training is entirely self-supervised. Deployed with monocular camera only.
How do we do it?Deep Distraction Suppression
[Barnes et al, “Driven to Distraction: Self-Supervised Distractor Learning for Robust Monocular Visual Odometry in Urban Environments”, ICRA 2018]
Generating Training DataDeep Distraction Suppression
[Barnes et al, “Driven to Distraction: Self-Supervised Distractor Learning for Robust Monocular Visual Odometry in Urban Environments”, ICRA 2018]
DeploymentDeep Distraction Suppression
[Barnes et al, “Driven to Distraction: Self-Supervised Distractor Learning for Robust Monocular Visual Odometry in Urban Environments”, ICRA 2018]
DeploymentDeep Distraction Suppression
[Barnes et al, “Driven to Distraction: Self-Supervised Distractor Learning for Robust Monocular Visual Odometry in Urban Environments”, ICRA 2018]
Deep Inverse Sensor Modelling In RadarRob Weston, Sarah Cen, Paul Newman, Ingmar Posner
[Weston et al, “Probably Unknown: Deep Inverse Sensor Modelling In Radar”, ICRA 2019]
Using Radar for Autonomous VehiclesGhost Reflections
Saturation
Cars
Speckle Noise
Free
Occluded
Noise Artefacts
Occlusion ?
Why? •Resilient to adverse weather conditions •Large range
Challenges? •Sensor Noise •Occlusion
[Weston et al, “Probably Unknown: Deep Inverse Sensor Modelling In Radar”, ICRA 2019]
Project Proposal
Aim: Leverage the power of deep learning to determine what space in the world is occupied in radar
Challenges: Sensor Noise, Occlusion, Training Labels
Solution: Assume radar is inherently uncertain to account for sensor noise and train using partial labels generated in laser
[Weston et al, “Probably Unknown: Deep Inverse Sensor Modelling In Radar”, ICRA 2019]
Our Approach
Use a deep Neural Network Model Heteroscedastic Aleatoric Uncertainty
Train using partial lidar labels
Desired Probability of Occupancy Grid
Raw Radar Input Lidar occupancy labels
?
[Weston et al, “Probably Unknown: Deep Inverse Sensor Modelling In Radar”, ICRA 2019]
TrainingGenerate lidar labels
Free
Occupied
Partially Observed
Unobserved
Observed
[Weston et al, “Probably Unknown: Deep Inverse Sensor Modelling In Radar”, ICRA 2019]
Uncertainty: Knowing what we don’t know…
[Weston et al, “Probably Unknown: Deep Inverse Sensor Modelling In Radar”, ICRA 2019]
Results ‣ Our model successfully outperforms classical CFAR filtering approaches
Raw Radar
Ground Truth
Our Approach
2D cartesian CFAR
1D polar CFAR
[Weston et al, “Probably Unknown: Deep Inverse Sensor Modelling In Radar”, ICRA 2019]
Use case: mapping
[Weston et al, “Probably Unknown: Deep Inverse Sensor Modelling In Radar”, ICRA 2019]
• machine learning does fit in (but maybe not in ways we first thought).
• systems engineering remains at the heart of deployable autonomous driving (at least for now).
• leverage the system (by definition it is all-knowing)
• be clever about how you use your data
Take aways for deployable driving solutions…
Maximum Entropy Deep IRL
[Wulfmeier et al, “Maximum Entropy Deep Inverse Reinforcement Learning”, arXiv, July 2015][Wulfmeier et al, “Watch This: Scalable Cost-Function Learning for Path Planning in Urban Environements”, IROS 2016]
• First to frame Max. Ent. IRL in deep learning context
• Efficient large-scale learning from demonstration… • backing out observed reward structure • refining human prior
Learning to act from lots of demonstration
Learned where to drive from 25,000 real-world trajectories.
SQAIRGenerative modelling of moving objects
[Kosiorek, Kim, Posner, Teh, “Sequential Attend, Infer, Repeat: Generative Modelling of Moving Objects”, NeurIPS 2018]
• Extends AIR to image sequences.
• Learns to detect and track objects without any supervision.
• Learns object motion models.
• Obtains a richer generative model by leveraging temporal information.
Code available on GitHub
…is mobile autonomy