Alexis.Pelletier
Capstone project · CTA

Robot Learning for Aerospace

Aerospace builds many variants in small batches, which makes conventional robot programming too slow. My pipeline teaches a robot a task from a simple human video.

Role
Design and development of the software pipeline
Context
MEC8370, Team 12, for the Aerospace Technology Centre (CTA)
Year
2026 – 2027
Tools
Isaac Sim, Isaac Lab, LeRobot, ACT, SmolVLA, SAM 2, MediaPipe, Python
44 of 50 trialssucceeded in simulation on a real task filmed at the CTA (88%, 95% CI: 76 to 94%).

From video to robot

The key idea: a robot doesn't have a human body, so it shouldn't copy human motions. It should achieve the same result on the objects.

Measuring honestly

Success rate over 50 identical trials

Standing-profile placement task, in simulation

Scripted expert
100 %
ACT, 500 demos
88 %
ACT, 198 demos
80 %
SmolVLA
78 %
ACT, MimicGen
50 %

The gap between ACT and SmolVLA is not statistically significant (McNemar, p = 0.27). ACT does predict 12 times faster: 34 ms versus 418 ms, measured on the cube task.

What I took away

  1. Demonstration quality matters more than quantity: 198 expert demos perform almost as well as 500.
  2. One physics setting can change everything: fixing gripper torque took grasps from 5% to 100%.
  3. A lax success criterion counts a falling part as a success. It has to require stability.
  4. A newer model is not better by default. It has to be proven with paired trials.

Project status: the chain works end to end in simulation. Transfer to the real robot is waiting on hardware.