problem of simulate the ego car

[ZOUYIyi]

I read the code to find when simulate the ego car in one policy in function EgoAgentForwardSim ,and it uses the function PropagateOnceAdvancedLK,it seems like that it ignores the ego.lon_behavior ,then just one step forward with some caculated velocity.

[MasterIzumi]

@ZOUYIyi In our implementation, the longitudinal behaviors are realized using different parameter sets of the longitudinal controller (i.e., IDM controller). In UpdateSimSetupForScenario, for each action (e.g., DcpLonAction::kAccelerate), we determine a set of IDM parameters for the forward simulation. Different settings of the longitudinal controller can generate different simulated trajectories.

[ZOUYIyi]

@MasterIzumi thanks ! I find that , I will read it more carefully...

[ZOUYIyi]

I read the code for simulate the npc cars , it looks like that they just go along with the ref_lane ,and the property : lat_probs is not used. so how can we use the uncertainty of this lat_probs

[MasterIzumi]

@ZOUYIyi Hi, sorry, the released implementation is not complete yet. As stated in the paper, the 'Guided Observation Branching' part handles the intention prediction uncertainty, and these functions will be realized in SimulateActionSequence. It will firstly conduct a preliminary safety assessment and generate several risky 'scenarios', then launch multiple 'sub-threads' for forward simulations. Currently, the released version only considers the 'most likely scenario', i.e., only one scenario for each action sequence. We will clean the code and release these parts in the future.

[ZOUYIyi]

@MasterIzumi Thanks still . Is the core/behaivor_planner a complete work for MPDM algorithm ? Because it seems like a substitute for eudm algorithm. If it is complete , it is still a good choice.

[MasterIzumi]

@ZOUYIyi Yes, this is our implementation of the basic version of MPDM (from the paper - MPDM: Multipolicy decision-making in dynamic, uncertain environments for autonomous driving), which also uses the most likely intentions in the forward simulation. The full version of MPDM (Multipolicy decision-making for autonomous driving via changepoint-based behavior prediction: Theory and experiment) that contains a changepoint-based prediction is not involved in this package.

[ZOUYIyi]

@MasterIzumi Thanks for your reply