Gasgoo Awards 2022 applicant: Long-short term Decoupled Decision-Making and Planning System from Zhito Technology

For the Gasgoo Awards 2022, Long-short term Decoupled Decision-Making and Planning System from Zhito Technology has applied for the Top 100 Players of China's New Automotive Supply Chain. 

Photo  credit: Zhito Technology

Product: Long-short term Decoupled Decision-Making and Planning System

Description

Instead of designing an integrated component of decision-making and trajectory planning as in conventional approaches, we decouple them into independent modules. The reinforcement learning agent in the decision-making component approximates the inference of future possible decision space, allowing the long-term, complex, uncertain, and interactive reasoning of ego behavior. The decision module outputs expected semantic action to the trajectory planning component, such as following, lane change, overtake， etc. The trajectory planning module uses an accurate kinematic model and object prediction, focusing on planning short-term, time-critical, collision-free, and comfortable trajectories, to achieve the received semantic action.

Photo  credit: Zhito Technology

Unique advantages

Both decision-making and trajectory planning have advantages and drawbacks in their approaches. But each of their pros and cons can be complementary. The inference of the environment model enables applying flexible methods in decision-making to address complicated problems efficiently. However, the resulting decision can be coarse and error-prone. On the contrary, accurate continuous modeling guarantees the accuracy, safety, and comfort of trajectory planning, but this also limits the pool of methods to choose from and the scale of the problem to solve. One way to utilize the pros and avoid the cons is to decouple decision and planning, by having decision making to decide on a semantic action under a large-scale complex environment, and letting trajectory planning to focus on achieving such specific action and avoid possible collisions. This is intuitively similar to human driving behavior, where we partition the driving scenario into time-spatial spaces, estimate the sequence of spaces to pass, then plan for throttle and steering within such spaces. 

Application

In busy traffic, where one has to make long-term lane-change planning to pass through slow vehicles; or in an unprotected left turn where you have to interact with oncoming vehicles. In these scenarios, decoupled decision making and trajectory planning free the decision component from meticulous trajectory optimization, with better capability to make long-term plans. The trajectory planning module makes sure the resulting behavior action can be achieved to the best as long as it is safe. 

Prospect

With more applications of automated driving systems in more complex scenarios, it is becoming more and more demanding for autopilot to address realistic situations. More new and powerful algorithms are emerging. However, they can never be free from the curse of dimensionality. Hence, decoupled decision and planning is a very promising framework to satisfy the spiraling demand for computational power.