End-to-end Deep Models For Self-driving Car

Transcript

1. End-to-end Deep Models For Self-driving Car Yuchu Luo Intelligent Agent Group 26 November, 2017
2. Rule based End-to-end Categories of driving systems
3. Huval, Brody, et al. "An empirical evaluation of deep learning on highway driving." arXiv preprint arXiv:1504.01716 (2015). Rule based Ulbrich, Simon, et al. "Towards a Functional System Architecture for Automated Vehicles." arXiv preprint arXiv:1703.08557 (2017).
4. Rule based 系统复杂度⾼ 依赖⾼精度地图 计算资源需求⾼ ⽆法适⽤所有情况 Road-level modeling Navigation Lan-level localization Feature-level localization Road-level localization Context modeling Feature extraction Guidance Stabilization
5. Rule based End-to-end 功能 Reactive control （边打电话边开车） ✅ ✅ Proactive planning (思考判断怎么开) ✅ ❌ 系统复杂度 极⾼ 极低 可解释性 ⾼ 低 ⼴铺成本 ⾼（HD Map） 低 传感器成本 极⾼ 低 车载计算能⼒要求 极⾼ 低 Rule based vs End-to-end From Baidu Apollo open course
6. End-to-end Deep Neural Network Steering Mariusz et .al 2016 Image
7. End-to-end ALVINN - 1989 Pomerleau, Dean A. "Alvinn: An autonomous land vehicle in a neural network." Advances in neural information processing systems. 1989.
8. Muller, Urs, et al. "Off-road obstacle avoidance through end-to-end learning. “ Advances in neural information processing systems. 2006.APA Figure: The steering angle produced by the system (black) compared to the steering angle provided by the human operator (red line) for 8000 frames from the test set. Very few obstacles would not have been avoided by the system. End-to-end DAVE - 2006
9. DeepDriving - 2015 Chen, Chenyi, et al. "Deepdriving: Learning affordance for direct perception in autonomous driving." Proceedings of the IEEE International Conference on Computer Vision. 2015. End-to-end (Intermediate Approach)
10. DeepDriving - 2015 dist_LL dist_MM dist_RR 1) angle: angle between the car’s heading and the tangent of the road “in lane system”, when driving in the lane: 2) toMarking LL: distance to the left lane marking of the left lane 3) toMarking ML: distance to the left lane marking of the current lane 4) toMarking MR: distance to the right lane marking of the current lane 5) toMarking RR: distance to the right lane marking of the right lane 6) dist LL: distance to the preceding car in the left lane 7) ··· Learn the traffic representation Chen, Chenyi, et al. "Deepdriving: Learning affordance for direct perception in autonomous driving." Proceedings of the IEEE International Conference on Computer Vision. 2015. End-to-end (Intermediate Approach)
11. End-to-end (Intermediate Approach) DeepDriving - 2015 Chen, Chenyi, et al. "Deepdriving: Learning affordance for direct perception in autonomous driving." Proceedings of the IEEE International Conference on Computer Vision. 2015. The ConvNet processes the TORCS image and estimates 13 indicators for driving. Based on the indicators and the current speed of the car, a controller computes the driving commands which will be sent back to TORCS to drive the host car in it.
12. End-to-end (Intermediate Approach) - 拥有⼀定的可解释性 - 较少的⼈⼯设计模块 - 难以设计⾜够的特征应付所有复杂情况
13. End-to-end Self-driving Car - 2016 Bojarski, Mariusz, et al. "End to end learning for self-driving cars." arXiv preprint arXiv:1604.07316 (2016). 仿真实验中，90% 的情况下 CNN 可以自动驾驶 路测表明，不同路况自动驾驶的概率为 98%
14. Codevilla, Felipe, et al. "End-to-end driving via conditional imitation learning." arXiv preprint arXiv:1710.02410 (2017). End-to-end Conditional Imitation Learning - 2017 The vehicle was given the command “turn right at the next intersection”
15. Dataset 真实数据 模拟器数据 • Udacity • Oxford • Comma.ai • Berkeley • Baidu Apollo • OpenAI • Universal • DeepMind • TORCS • Virtual KITTI
16. Berkeley Data Drive Over 400 hours of HD video sequences across many different times in the day, weather conditions, and driving scenarios. Our video sequences also include GPS locations, IMU data, and timestamps. 30 fps 720 p
17. 如何将这些数据集 用于真实的无人驾驶任务？ Imitation Learning - Behavior Cloning CNN b b (state_0, action_0, state_1, action_1, …, state_n) state: images from 3 camera action: steering angle
18. Self-driving Formulation - Egomotion Prediction α s F(s,a):S × A → ! • State s 当前状态（包括历史信息） • 视觉信息 • 汽车动⼒状态 • Action a: • 离散： GO, STOP, LEFT, RIGHT • 连续：下 0.1s 的转弯⾓度和加速度 • 6 DOF (⾃由度) motions 普适性强，不受汽车 自身参数影响
19. Problem: 复合误差（Compounding errors） Distribution mismatch
20. Self-driving Car - 2016 Bojarski, Mariusz, et al. "End to end learning for self-driving cars." arXiv preprint arXiv:1604.07316 (2016). 仿真实验中，90% 的情况下 CNN 可以自动驾驶 路测表明，不同路况自动驾驶的概率为 98% How?
21. Self-driving Car - 2016 Bojarski, Mariusz, et al. "End to end learning for self-driving cars." arXiv preprint arXiv:1604.07316 (2016). self-stability 左右摄像头可以令无人车 从部分糟糕的状态中恢复
22. The better way… π: policy make P data (O t ) = Pπ (O t )
23. Bojarski, Mariusz, et al. "End to end learning for self-driving cars." arXiv preprint arXiv:1604.07316 (2016). Experiments
24. How to model historical information How are you <EOL> W I am fine <EOL> Seq2Seq 模型（RNN）
25. RNN based architecture
26. Xu, Huazhe, et al. "End-to-end learning of driving models from large-scale video datasets." arXiv preprint arXiv:1612.01079(2016).
27. Discrete Action Driving Model Xu, Huazhe, et al. "End-to-end learning of driving models from large-scale video datasets." arXiv preprint arXiv: 1612.01079(2016).
28. Continuous Action Driving Model Xu, Huazhe, et al. "End-to-end learning of driving models from large-scale video datasets." arXiv preprint arXiv: 1612.01079(2016).
29. 其他值得做的点： • Attention mechanism • 横纵向关联关系 • 时序处理
30. Conditional Imitation Learning - 2017 Codevilla, Felipe, et al. "End-to-end driving via conditional imitation learning." arXiv preprint arXiv:1710.02410 (2017). High-level overview
31. Conditional Imitation Learning - 2017 Network architecture I: command input
32. Conditional Imitation Learning - 2017 Network architecture II: branched
33. Tactical decision making for lane changing with deep reinforcement learning Mukadam, Mustafa, et al. “Tactical Decision Making for Lane Changing with Deep Reinforcement Learning." (2017).
34. Framework
35. Our work
36. Task 1: from real to virtual Virtual Data Real Data Virtual Env Real Env Deep Reinforcement Learning Generative Adversarial Network
39. Virtual Data Real Data g(x) f(x) Reward Mean square error DNN adversarial loss
40. Task 2: conditional imitation learning for real cars Furthermore… Use reinforcement learning for high-level control conditional imitation learning for low-level control
41. VIDEO
44. Goal: give a GPS coordinate, then get there