Research | Mobile Robotics Lab @ McGill University

Theme 01

Reinforcement Learning

MRL develops reinforcement learning methods for continuous control and offline decision making, including widely used algorithms for reducing value error, learning from fixed datasets, and improving online and offline control.

TD3 BCQ TD3+BC TD7

Theme 02

Imitation Learning

We develop imitation learning methods that let robots acquire skills from demonstrations or observations, with stable and contractive policies for safer recovery outside expert trajectories.

Optimal Transport IL Stable Neural Policies Contractive Imitation Contractive Diffusion

Morphology-conditioned quadrupedal world model

Theme 03

World Modeling for Robotics

We study world models that capture robot dynamics for planning, search, and behavior learning, including morphology-conditioned models for quadrupeds and model-based RL work on when search helps or hurts decision-making.

QWM Search in Model-Based RL

Theme 04

Field Robotics

Field robotics at MRL is an umbrella for deployed autonomy across forests, reefs, reservoirs, and other unstructured environments. These projects cover mapping, traversability, off-road planning, aerial marine localization, underwater autonomy, and surface vehicles.

Topological exploration

Traversability-aware maps for long-range off-trail navigation.

Vision-based traversability

Visual and multimodal terrain assessment for off-road autonomy.

Off-road attention

Trajectory-constrained visual attention for local planning.

Underwater / Aqua

Underwater navigation, visual policies, and reef autonomy.

Aerial marine localization

Drone-based localization for surface and near-surface robots.

On-water / boats

Surface vehicles for probing, sampling, and monitoring.

Topological Off-road attention Underwater Aerial localization Surface vehicles

Theme 05

Vision-Language Models for Robot Learning

We explore how pre-trained vision and language models can ground semantic instructions, identify useful visual structure, and support robot policies, with current work on tactile VLA fusion and foundation-model reasoning for safer robot behavior.

TacFiLM

Vision-language models for robot learning research

Theme 06

Dexterous Manipulation

We study contact-rich manipulation with multi-fingered hands and robot arms. Current work includes DexSuite, a unified simulation and benchmarking framework for dexterous manipulation, and the LEAP Hand, a low-cost anthropomorphic hand for robot learning.

DexSuite