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%
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.
We believe that whether it’s a World Model [9], a VLA model [11], or an LLM controlling the robot [8] [10], it will inevitably require a reliability-improvement phase in the form of online learning.
History shows that the key to turning GPT-3 into ChatGPT was the RLHF process. DeepSeek R1 achieved strong reasoning behavior through RL on verifiable tasks. Claude Code is impossible to imagine without the intense use of RL. A humanoid can learn to stand and walk through RL training in simulation, even without any prior pretraining.
What all of these have in common is that they used real model rollouts, and the model learned “online” from the experience gained during those rollouts.
Let’s look at the tools currently used to improve reliability in robotic manipulation and why they don’t work.
Unfortunately, the tool that works well for other applications doesn’t work here. Simulation and world models have a significant gap from reality for manipulation tasks. RL on a real robot also damages materials and the robot itself during the exploration phase. RL is highly sample inefficient by design. It is like trying to sucking supervision bits through a straw. It would take 100 years to collect 1 million one-minute trajectories on a real robot. Creating the reward function is also a challenge. Often, you do not just want the task to be just completed, but to be completed with clean motion.
SFT / UMI gripper [5] / imitation learning