Continual learning is the process by which an AI model continuously improves and adapts based on new experience, rather than being trained once and then remaining fixed. [6] We got a taste of it and want to share it with you.
On a benchmark task, our pipeline required ~10x less data and operated at around 100% reliability, working autonomously for 45 mins in a row, compared to ~85% for the next best model.
https://www.youtube.com/watch?v=C22W0AHP4NA
Zip tie benchmark task, comparing with the next best AI pipeline for precise manipulation.
| Ours | Pi06+RLT [1] |
|---|---|
| Success rate: 97% | Success rate: 85-90% |
| Training data: 29 mins | Training data: 5 hours + 500 RL episodes. |
| Throughput: 23 ties per 10 mins | Throughput: 13 ties per 10 mins |
Success rate breakdown:
- Inserting: 100 out of 100 (100%)
- Flipping to the right side: 100 out of 100 (100%)
- Grabbing the zip-tie tail: 100 out of 100 (100%)
- Picking from the bin: 100 out of 103 (97%)
- Total: 97%
We use two parts in our AI pipeline.
One part is for directly performing the work in production. It is optimized for speed of execution and adapting to new examples for continuous learning. The task is divided into segments, which can be retrained separately.
We start with a very small amount of data, doing most of the data collection on-policy while the model is running. We use a new-state detection mechanism to stop the robot, add handling for that state to the dataset, and retrain the segment in 20 minutes.
The second part is for data collection.
An advanced coding LLM (Claude Code) decides how to split the task and determines how many episodes need to be collected for each segment.
It asks the human operator to place target objects in the required positions and moves the arm by adjusting waypoints to record examples. It makes several attempts to record a clean example, which then goes into training for the segment in the fast part.
Learning from experience has remained a coveted goal for roboticists for many years.
The building blocks of the process are known and described in papers, but making it real comes down to the difficulty of practical execution, which leaves room for startups with relevant experience to achieve better results than well-funded labs with purely academic expertise.
The difference between an 85% success rate and a 99% success rate is what separates a demo from a product.