Robotics research

LeRobot, SmolVLA, and SO-101 Evaluation

Research tooling and model evaluation for robotic-arm autonomy, with a focus on making training, deployment, and experiment review less repetitive and more measurable.

View Lab Tools

Evidence

Project media

Pick-and-place deployment clip from the robotic-arm policy rollout.

Two robotic arms mounted near a lab table for leader/follower robotics experiments. — Robotic-arm lab setup used for leader/follower data collection and policy deployment experiments.

Diagram showing allowed training positions for the pick-and-place block task. — Training-position diagram used to constrain the pick-and-place setup.

Training summary slide comparing SmolVLA batch sizes, deployment results, and training loss curves. — Training summary comparing batch-size conditions, rollout notes, and training-loss curves.

What I built

Lab workflow tooling

I built a Visualizer inside a LeRobot GUI Wrapper so the lab could reduce repetitive command generation and inspect datasets, deployment videos, model folders, metadata, history, and experiment runs from one interface.

The wrapper sits on top of an existing LeRobot installation instead of replacing it. It helps with hardware bring-up, teleoperation, dataset recording, training, deployment, replay, history review, and experiment comparison.

Model work

Training and evaluation

I trained and evaluated multiple robotic policies for pick-and-place tasks on SO-101 arms. The model families I worked with include SmolVLA, ACT, GR00T, and PI0.5-style workflows.

Training ran on an NVIDIA A100 cluster, and the research work focused on developing environments and training conditions that improved real hardware performance instead of only producing offline metrics. Early SmolVLA experiments showed that lower offline loss did not directly translate to better deployment; the 16-batch run had the strongest chunk deployment result in the initial comparison.

Hardware

SO-101 robotic arms

The work used 6-DOF SO-101 leader/follower robotic-arm setups for data collection and policy deployment. That meant dealing with calibration, camera setup, action chunking, dataset quality, and the difference between a model that trains and a model that completes the task on hardware.

Batch comparison

SmolVLA deployment notes

Condition	Batch	Steps	Loss	Gradient	Chunk result
Small batch	16	20,000	0.00468	0.13	5/10
Medium batch	32	20,000	0.0036	0.12	4/10
Large batch	64	20,000	0.0033	0.09	3/10

The first useful deployment note was practical: the model could recover well and pick consistently when the object started in the right region, but depth perception and approach angle were recurring failure modes.

This table is included as an example of the experiment-tracking format and preliminary deployment notes. Later runs produced better results; these numbers are not meant to represent the final outcome of the research.