Actions used to control agent rollouts

Container for sequences of 2D positions, yaws, controls, availabilities.

Upscaling then double conv

Guide trajectory to satisfy rules by directly perturbing it

A wrapper policy that can (optionally) control both ego and other agents in a scene

A multi-output MLP network: The model split and reshapes the output layer to the desired output shapes

MLP encodes past state history. Assumes inputs will be (x,y,hx,hy,s,l,w,avail)

Base class for simple Multi-Layer Perceptrons.

Log trajectories and other essential info during rollout for visualization and evaluation

Conv1d --> GroupNorm --> Mish

Metrics that report failures caused by either collision or offroad

Encoder is made up of self-attn and feed forward

A wrapper policy that feeds subgoal from a planner to a controller

Construct a layernorm module

Encodes map, may output a global feature, feature map, or both.

MLP encodes past state history.

Predict a spatial map same size as the input rasterized map

Map the encoded tensor to a description of the whole scene, e.g., the likelihood of certain modes

Container for sequences of 2D positions and yaws

Ego-centric model that is aware of other agents' future trajectories through auxiliary prediction task

Handles initializing guidance functions and computing gradients at test-time.

(convolution => [BN] => ReLU) * 2

empirical moving average

Core encoder is a stack of N layers

A wrapper policy that feeds plan samples from a stochastic planner to a controller

Vectorized Scene Component + Rasterized Map

A multi-input, multi-output MLP: The model flattens and concatenate the input before feeding into an MLP

Add Ornstein-Uhlenbeck noise to the trajectory

A wrapper that computes action yaw from action positions

Implement the PE function.

extension of the PE function, works for N dimensional position input

Implements FFN equation.

A basic image-based rasterized map encoder

A residual connection followed by a layer norm. Note for code simplicity the norm is first as opposed to last.

map the scene information to attributes that summarizes the scene

Note: only support batch size 1 for now. 1.takes in agent-centric data and convert it to scene-centric data 2.apply policy get_action

Agents should not collide with each other. NOTE: this assumes full control over the scene.