Posts
Notes from a small-budget robot-learning lab: hardware plans, VLA experiments, datasets, and failed runs.
-
Does 2.5× More Data Help a Robot Generalize? ACT vs. SmolVLA vs. π0.5
In my [previous experiment](https://merry-and-pippin.net/posts/can-a-robot-learn-81-pick-and-place-routes-from-only-25/), I tested whether a robot could learn 81 possible pick-and-place routes from just 25 demonstrated combinations.
-
Can a Robot Learn 81 Pick-and-Place Routes from Only 25?
Testing whether ACT, SmolVLA, and π0.5 can generalize to novel cube-to-cup routes after training on only 25 of 81 start-goal combinations.
-
Teaching Merry and Pippin a New Trick
Training an ACT imitation-learning policy for 20,000 steps and turning SO-101 demonstrations into autonomous movement with LeRobot.
-
The Commands Behind My LeRobot Workflow
The command-line workflow I use to collect LeRobot datasets, train ACT policies, and run ACT or SmolVLA inference on my two-camera SO-101 setup.