Can a robot learn a driving style from a few human demos?
Adding a small imitation-learning module between a classical path planner and motion controller let a robot copy a human's navigation style while still reliably reaching its goal.
Source
Improving Autonomous Robotic Navigation Using Imitation Learning
Study at a glance
- Design
- Computational / modelling — Simulated wheeled robot in Gazebo and Unity; a small CNN picks one of nine local waypoints; 50 trials per model per navigation task in training and unseen test environments.
- N
- No single N: 200 planner demonstrations and 12 human demonstrations per training environment; evaluation used six navigation tasks per simulator with 50 trials each.
- Population
- Simulated Clearpath Husky robot in two Gazebo and two Unity environments
- Outcome
- Navigation success rate and similarity (inverse modified Hausdorff distance) of trajectories to held-out human demonstrations
Structured fields used in claim comparison tables when every cited study has a complete layer.
What they did
The authors kept a standard robot navigation stack but inserted a learned module between the global path planner and the local controller; it chooses one of nine nearby waypoints from lidar, semantic camera features, and the goal's bearing and distance. The module was first trained by behavioural cloning on 200 runs of the classical planner, then fine-tuned on just 12 human tele-operated runs that followed a rule: give objects with a particular label a wide berth. It was tested in simulation against two published learning-from-demonstration approaches, in both training and unseen test environments.
What they found
The hybrid method kept a high success rate in all four environments while the baseline methods did not, and its paths were more similar to held-out human demonstrations. Differences were statistically significant in every environment except the simplest (the Gazebo training room). One baseline tended to wander around avoiding labelled obstacles without ever reaching the goal. Most failures of the proposed method came from choosing waypoints inside obstacle safety margins, leading to collisions.
The limits
What it doesn't show
Everything was done in simulation with a single demonstrator (one of the authors) and one simple style rule, so transfer to real robots, other people or richer preferences is untested. The paper reports results mainly in figures, with few exact success percentages in the text, and did not compare against the classical planner alone. Choosing from nine fixed waypoints limits how precise the learned motion can be.
Key terms
- Behavioural cloning
- Supervised learning that maps observed states to the actions a demonstrator took in them.
- Learning from demonstration
- Teaching a robot a behaviour by showing examples instead of programming rules or cost functions.
- Distribution shift
- When a cloned policy drifts into states the demonstrator never visited, so its training data no longer covers what it sees.
- Local planner
- The part of a navigation stack that turns a nearby goal into feasible wheel commands that respect the robot's kinematics.
- Catastrophic forgetting
- When further training on new data erases what a network learned earlier.
Flashcards
0 of 10 answers reviewed
Research intelligence for this paper
See its role on concept claims, tensions it is part of, placement history, and related discoveries.
Quiz yourself
What does the learned module output?
Common questions
Why not learn the whole controller end to end?
End-to-end policies need lots of data and must relearn obstacle avoidance and kinematics. Keeping the classical planner and controller lets a small module learn only the style from a handful of demos.
How did they fight distribution shift without more demonstrations?
They synthetically rotated each recorded state by plus or minus 45 degrees and labelled it with a corrective action, tripling the training data.
Why pre-train on planner runs first?
The planner cheaply provides many examples of basic goal-reaching, so the few human demos only need to teach the preference.
More on Robot learning