LePancake

we didn't have an idea and ashray was hungry

Comment

Inspiration

During the hackathon, one of our team members couldn’t stop thinking about breakfast. Fueled by late‑night coding and an intense pancake craving, we realized there was a perfect intersection between robotics and cooking—why not build a pancake‑flipping robot? From that moment on, LePancake’s mission was born: automate the most critical step in pancake making so nobody has to suffer the “burnt bottom” blues ever again.

What it does

LePancake uses a camera mounted above a griddle to detect the exact moment a pancake is ready to flip. Our vision model tracks each pancake’s edges and surface texture in real time, and as soon as it reaches the golden‑brown threshold, the robotic arm swoops in, flips it cleanly, and places it back down—every time, no soggy middles or scorching. With a single button press, you can queue up up to four pancakes, and LePancake handles the rest.

How we built it

  • Phospho & LeRobot: We leveraged Phospho’s leader-follower controller in order to easily control our robot when collecting training data and LeRobot's premade training scripts allowed for us to easily train on Google Colab without any fuss.
  • Hugging Face & PyTorch: Starting from a pre‑trained ResNet backbone, we fine‑tuned on our custom pancake‑flip dataset using PyTorch, then exported the model via the Hugging Face Hub for easy versioning and deployment.

Challenges we ran into

  • Dataset quality: Capturing enough variation in pancake shapes, sizes, and lighting conditions was harder than expected. Early models mis‑classified over‑cooked edges as “ready to flip.” We spent hours refining our data‑collection rig—adjusting griddle heat settings, repositioning lights, and diversifying pancake batters—to boost generalization.
  • Mechanical tolerance: Slight warping in the 3D‑printed arms led to inconsistent flips. We added washers to increase preload on certain joints to remove the inconsistency in our flips.

Accomplishments that we’re proud of

  • 92% flip success rate: Over hundreds of trials, LePancake gracefully flipped pancakes without breaking or over‑rotating 92% of the time.

What we learned

  • Data is king: Even the best architecture can’t compensate for a narrow or biased dataset. Investing in diverse, well‑labeled training data paid off faster than chasing fancy model tweaks.
  • Interdisciplinary synergy: Bringing together expertise in robotics, computer vision, and embedded systems exposed us to real‑world trade‑offs—latency vs. accuracy, robustness vs. complexity.
  • Rapid iteration matters: Prototyping in software-in-the-loop simulation first using IsaacSim and then switching to physical tests only when confidence was high saved us precious hackathon hours.

What’s next for LePancake

  1. Pouring & batter control: Automate the batter pour with a servo‑driven dispenser, enabling perfectly uniform pancakes without human intervention.
  2. Plating & topping: Add a second arm or conveyor to move flipped pancakes onto a plate and even drizzle syrup or add fruit.
  3. Model refinement: Expand our flip‑ready dataset with more griddle surfaces (nonstick, cast iron) and lighting conditions (daylight, dim kitchens) to push accuracy above 95%.
  4. User interface: Build a smartphone app or voice skill so you can say, “Alexa, make me pancakes,” and let LePancake handle breakfast from start to finish.

Built With

Share this project:

Updates