About
What is VRAI?
VRAI (Vanier Research in Artificial Intelligence) is a student-run research lab at Vanier College. We work on embodied AI — teaching physical robots to perceive, reason, and act in the real world.
What we work on
Imitation Learning
We teach robots by recording human demonstrations — the robot watches, learns the underlying behaviour, and replicates it autonomously. Our primary framework is Hugging Face's LeRobot.
Computer Vision
Our robots perceive the world through multi-camera arrays. We process those visual streams with convolutional networks to give the model a rich understanding of its environment.
Edge Computing
Inference runs locally on NVIDIA Jetson hardware, making our platforms fully self-contained. No cloud dependency — the robot thinks for itself, onboard.
Our pipeline
01
Build & Calibrate
Assemble the robot, flash the compute, and calibrate every joint and camera until teleoperation is clean and reliable.
02
Collect Data
A human operator controls a leader arm while the follower records — capturing joint states and camera feeds as a high-quality dataset.
03
Train the Model
We feed the dataset into an imitation learning model (currently ACT) and train until the robot can reproduce the demonstrated behaviour.
04
Run Inference
The trained model runs onboard the Jetson. The robot operates autonomously, using only its cameras and joint sensors.
Hardware we use
- Hugging Face LeRobot SO-101 dual-arm setup (leader + follower)
- NVIDIA Jetson (Jetpack 6.2) for onboard inference and teleoperation
- Multi-camera arrays (side view + wrist/grip view)
- Yahboom ROSMASTER X3 Plus (archived)
Frameworks & tools
- Hugging Face LeRobot — dataset collection, training, and inference
- ACT (Action Chunking Transformer) — our primary imitation learning model
- ROS — used during the Yahboom phase for sensor integration
- PyTorch — model training and experimentation
Curious how the models actually work?
We wrote deep dives on the architectures we use.