Computational Teaching for Driving via Multi-Task Imitation Learning

TL;DR

The experiments show that the right set of auxiliary machine learning tasks improves prediction of teaching instructions, and students exposed to the instructions from the teaching system improve their ability to stay within track limits, and show favorable perception of the model's interaction with them, in terms of usefulness and satisfaction.

Abstract

Learning motor skills for sports or performance driving is often done with professional instruction from expert human teachers, whose availability is limited. Our goal is to enable automated teaching via a learned model that interacts with the student similar to a human teacher. However, training such automated teaching systems is limited by the availability of high-quality annotated datasets of expert teacher and student interactions that are difficult to collect at scale. To address this data scarcity problem, we propose an approach for training a coaching system for complex motor tasks such as high performance driving via a Multi-Task Imitation Learning (MTIL) paradigm. MTIL allows our model to learn robust representations by utilizing self-supervised training signals from more readily available non-interactive datasets of humans performing the task of interest. We validate our approach with (1) a semi-synthetic dataset created from real human driving trajectories, (2) a professional track driving instruction dataset, (3) a track-racing driving simulator human-subject study, and (4) a system demonstration on an instrumented car at a race track. Our experiments show that the right set of auxiliary machine learning tasks improves performance in predicting teaching instructions. Moreover, in the human subjects study, students exposed to the instructions from our teaching system improve their ability to stay within track limits, and show favorable perception of the model's interaction with them, in terms of usefulness and satisfaction.

Figures (3)

• Figure 1: Computational Teaching Model overview. Our model takes as input the student's past driving history as a sequence of states and controls along with local map information. The model outputs: (i) probability estimates for teacher actions (ii) future trajectories and (iii) skill estimates for the student. Our model learns to predict the correct teacher actions and can be utilized within an automated teaching system for track driving.
• Figure 2: Model architecture for computational teaching and multi-task learning. The inputs consist of a sequence of past trajectories, $(\tau^{-N:0})^P$ encoded via a trajectory encoder, and the corresponding local map representations $m_{local} \in \mathcal{M}_{local}$ encoded via a graph attention model, along with information about the optimal racing line for experiments with racing data. The trajectory encoder is an MLP+transformer Ngiam2022-ny for experiments with the semi-synthetic dataset and an MLP+LSTM alahi2016social for the professional track driving instruction dataset. The trajectory and map embedding are fused via a cross-attention mechanism to generated an encoded state. The latent embeddings are computed by applying a series of attention based and pooling operations and then get fed to three MLP decoders for the primary (teacher action) and auxiliary (skill and trajectory) prediction tasks.
• Figure 3: Evolution of objective metrics during the course of the study. (a) Normalized Lap time. (b): Normalized Percentage out of bounds. The dark gray colored band is the familiarization, the light gray correspond to the training, and the light yellow band to the retention phases.