MEGA-DAgger: Imitation Learning with Multiple Imperfect Experts

All research

Related publications

Overview

This project brings together two autonomous-racing studies: MEGA-DAgger, which learns from multiple imperfect experts, and a benchmark that evaluates imitation learning both on its own and as an initialization for reinforcement learning.

The descriptions below follow Xiatao Sun’s original project pages: MEGA-DAgger and Imitation Learning for Autonomous Racing.

Project video

Learning from multiple imperfect experts

Interactive imitation learning methods are often built around a single ideal expert, although practical systems may instead rely on several imperfect ones. MEGA-DAgger extends DAgger to this setting.

During dataset aggregation, it excludes unsafe examples. To handle disagreement about the action to take, it compares experts using criteria tailored to the racing scenario.

Two imperfect expert trajectories contain different unsafe segments; the learner combines their complementary good behavior.
Learning complementary behavior instead of copying every expert mistake. Figure 1 from the MEGA-DAgger paper.
MEGA-DAgger control loop: choose an expert, gate expert and novice actions, filter unsafe data, and resolve conflicts.
Expert selection, data filtering, and conflict resolution augment the interactive learning loop. Figure 2 from the paper.

The autonomous-racing experiments show that the resulting policy can surpass the individual experts and Human-Gated DAgger.

MEGA-DAgger: Xiatao Sun, Shuo Yang, Mingyan Zhou, Kunpeng Liu, and Rahul Mangharam.

Benchmarking imitation and reinforcement learning

The companion study tests how imitation-based policies perform in autonomous racing, and whether they provide a useful starting point for reinforcement learning. It evaluates scaled racing vehicles in simulation and in physical experiments.

Its findings favor interactive approaches over traditional imitation methods. Their more efficient use of demonstrations also helps improve subsequent reinforcement learning. The benchmark is intended to support further research combining these approaches for autonomous racing.

F1TENTH race cars driving through an indoor course in Levine Hall.
Physical F1TENTH evaluation in Levine Hall. Photograph from Xiatao Sun's autonomous-racing benchmark page.

Benchmark study, IEEE IV 2023: Xiatao Sun, Mingyan Zhou, Zhijun Zhuang, Shuo Yang, Johannes Betz, and Rahul Mangharam.

Demonstrations and presentation

An F1TENTH car follows another car on an outdoor track, with an onboard camera view inset.
Racing and onboard views from Xiatao Sun's MEGA-DAgger project page.
More videos and the original workshop presentation

MEGA-DAgger supplementary video

HG-DAgger and PPO on the physical car

First-person racing view

IROS 2022 MiniRobot Workshop slides

Workshop presentation recording

Papers and materials

Project content, photographs, and graphics are credited to Xiatao Sun and the paper authors. Xiatao carried out this work with xLAB during his time at Penn.