group05-vcr · Robotics Pizza Hack 2026

A robot arm that puts cheese and salami on a pizza, learned from human demonstrations. Team group05 (VCR) · Robotics Pizza Hackathon, 2026.

We teach an ElRobot arm (7 joints + gripper, driven by NormaCore Station) two skills by imitation learning: put cheese on pizza (scoop cheese from a bowl, sprinkle it over the pizza) and place salami on pizza (pick one slice from a tray, place it on the pizza). Humans teleoperate a leader arm; we record what the follower arm, two cameras and the joints do, and train policies that reproduce the movement.

Models

Start with 05-vcr-salami-smolvla-v7 for salami (our latest, and the most decent on the real robot) and 05-vcr-cheese-smolvla-v1 for cheese.

Every model takes two 224×224 camera pictures (image 0 = table camera, image 1 = arm/gripper camera), the 8 joint positions (normalised 0–1) and the task sentence, and returns a chunk of 50 target joint positions (10 Hz = 5 s).

Datasets

Common format: 10 Hz, two JPEG cameras (224×224, table first, arm second), 8 joints (position_norm = state, goal_norm = action), one task sentence per episode. All datasets share one calibration (including the gripper range 1327–2769) and the same rest pose, so a cheese stage and a salami stage can follow each other.

What we learned (honest results)

• Data quality mattered more than model choice. We found and removed mixed calibrations, glitchy frames, long still tails (59% of the frames near the rest pose were "do nothing"), and episodes that wait 10–15 s before moving.
• Starting from the exact rest pose is hard: the picture is identical whether the demo said "wait" or "go", so models flip or freeze. Starting from a ready pose above the tray works far better; our runner moves the arm there first.
• Cutting the pauses out of the demonstrations helped: v7 (same 33 episodes as v2, but with the still tails trimmed, the pauses cut to 0.2 s and image augmentation) is the most decent model we tested on the robot.
• Open problems: grab precision (few salami positions in the data) and hesitation above the food.
• Image augmentation lowered the loss on held-out episodes (0.95 vs 1.12 on 4 held-out episodes). We also tested Cosmos 3 Nano (video generation): it keeps motion with edge control, but generating all episodes was too slow for the time we had.

How we run it

Training: NormaCore SmolVLA trainer (and LeRobot for π0.5) on a Nebius GPU. Inference: a small GPU server (FastAPI) holds the models and answers "picture + joints + sentence → next 50 target positions" in about 0.2 s; a runner next to the robot sends the Station frames to it and plays the targets with safety limits (clamped steps, stop on network errors, hand-over from a scripted ready pose).

Team

• Ricards Liskovskis
• Vincent Kamusella
• Claus Nottbrock