【業界大牛】Mobileye on CES2107

Mobileye 定義

  • sensing
  • mapping
  • Driving Policy (30:00)
    • Merging with Human Traffic is difficult.
    • Reading their intention from other’s motion and planning cause huge processing load.
    • Polishing Driving Policy needs machine learning. (34:09)
      • Nowadays, the method lead to too simplistic policy.
        • Rule-based Approach
        • Brute-force Pruning all possibility
      • Multi-Agent Game
        • Intention “What’s the road user about to do?"
          • Reading their intention from other’s motion
        • Defensive / Aggressive Trade-off
          • Region dependent point on the spectrum.
        •  Negotiation
          • Signal intention to other driver through action
          • Need to balance “Safety Guarantee" and “blend in" dense traffic according to the “norms" of the region.
      • Uncertainty
        • sensing
        • disagreements between different sensors
        • uncertainty about road user’s intention
      • Pros:
        • It’s easier observe and collect data than understanding the underlying rule of the problem you want to solve.
      •  Cons:
        • ML is based on statistics of data and ability to sift(Scale-invariant feature transform) through data.
        • Thereby, falling on “corner" case.
          • Take long time to collect them.
          • Safety Concern
      • Key Algorithm Difference (37:00)
        • Sensing → Sense the present → Deep Supervised Learning
          • No effect on the environment
          • Data for training and validation can be collected in advance.
        • Driving Policy → Planning the future →  Reinforcement Learning
          • Action (or Prediction) Affect the environment
          • Online Data Collection : Data for training can not be collect in advance.
            • Version Upgrade need another round of data collection.
            • Become unwieldy to flush corner case.
          • 【Paper】Safe, Multi-Agent, Reinforcement Learning for Autonomous Driving
          • Three Criteria
            • Safety
            • Probability of the Maneuver
            • Processing Time
      • 【開大絕了】都用我的演算法吧 XD (43:26)
        • Data Sharing
        • Standardization for the Validation
  • L4 / 5 Fully Autonomous (44:00)
    •  Redundancy
    • Scalable Architecture

 

廣告
本篇發表於 產業動態。將永久鏈結加入書籤。

發表迴響

在下方填入你的資料或按右方圖示以社群網站登入:

WordPress.com 標誌

您的留言將使用 WordPress.com 帳號。 登出 /  變更 )

Google+ photo

您的留言將使用 Google+ 帳號。 登出 /  變更 )

Twitter picture

您的留言將使用 Twitter 帳號。 登出 /  變更 )

Facebook照片

您的留言將使用 Facebook 帳號。 登出 /  變更 )

連結到 %s