30+ teams locked in for 36 hours at the Physical AI hackathon we hosted last weekend in SF ππ€
We equipped them with lerobot arms to solve three tasks: shape insertion, charger plugging, and liquid pouring.
Here are the 7 winners and what they built (with github and video): π§΅π
π₯1st place: ETA0.1
They built an Autonomous Robot Barista using an ACT policy. βοΈ
95% accuracy and zero spills after ~100k steps. They introduced real-time perturbations (moving the cup mid-pour), and with just 10k fine-tuning steps, the robot learned to adapt on the fly.
Code: https://t.co/glKK1MrIfv
Team: @sangam_chapagai, @MahimanaB, @DhabariaAnjali, Gary Lim, Benedict Chan
π₯2nd place: Pivot
To address the combinatorial challenges of long-horizon robotic manipulation, Team PIVot implements a hierarchical architecture that pairs a low-frequency VLM "thinker" for high-level planning with a high-frequency, language-conditioned policy for skill execution.
Code: https://t.co/RvOB0wSshO
Team: @praveenvnktsh, @vib2810_, @shriishwaryaasv
π₯3rd place: Automate
Trained an ACT policy from scratch on a real robot and reached ~90% success on a multi-puzzle insertion task with minimal data, showing emergent recovery and alignment behaviors. A great reminder that high-quality demos beat pretraining when working with real-world robotics.
Github: https://t.co/7CWnmtVkJE
Team @Poorvi_rh, @anishmadan23, @SamuelPfrommer
ποΈWorld Intelligence Award: ServingU
A classic "divide and conquer" engineering approach. π₯€
They split pouring into 3 subtasks, training and benchmarking ACT vs. Diffusion for each (ACT won). Finally, they built a custom script to orchestrate the models sequentially, achieving pixel-perfect coffee pouring.
Code: https://t.co/y4ZmI6GSYR
Team: @neel7281@JesusBetan86866, @mark_mau_, Liu Cathy, Garima Bhandari
ποΈProtocol Labs Award: Ladybug
An autonomous physical audiobook reader. ππ
The arm turns the page, scans it (OCR), streams text-to-speech, and waits for the audio to finish before turning the page again. The loop: Manipulate β Read β Speak β Repeat.
Github: https://t.co/8IHqro26Ih
Team: Alison Cossette, Andreea Turcu, Sudhir Dadi
ποΈActiveloop Award: Robocafe
A voice-controlled butler. π£οΈπ¦Ύ
This team trained a robot to respond to voice prompts to serve croissants, pour water (half or full glass!), and even clean up the table afterwards. Complex task chaining handled by different ACT models.
Github: https://t.co/lr4ZjgRBXY
Team: Nestor Tkachenko, Khoi Le