One thing that stood out to me while going through the @continuumlabs_ GitHub is that collecting robot data isn’t the finish line. It’s actually where another problem begins. What do you do with all that raw data? Imagine recording thousands of robot demonstrations. Some movements might be too noisy. Some trajectories might have inconsistencies. Some data might not be useful for training at all. You can’t just throw everything into a model and hope for the best. This is where Continuum Data Cleaning comes into the picture. It’s a toolkit in the Axis ecosystem focused on processing and evaluating robot trajectories before they’re used further down the pipeline. And honestly, this part doesn’t get talked about enough. We hear a lot about collecting more data. But more data isn’t automatically better data. Think about a robot trajectory as a recording of everything that happened while completing a task. If the recording contains unnecessary noise or irregularities, the model may end up learning from those imperfections instead of learning the behavior you are actually interested in. So the data needs to be inspected and processed. Continuum Data Cleaning includes tools for things like trajectory smoothing and resampling, alongside metrics that can help evaluate the quality of trajectories. For example, smoothing can help reduce unwanted noise in the movement data. Resampling can make trajectories more consistent in how their data points are distributed over time. And evaluation metrics can give you a way to understand whether a trajectory looks reasonable instead of simply assuming that every recorded demonstration is useful. There is also task level validation. That’s important because a trajectory can look technically fine while still failing to accomplish the actual task. A robot moving smoothly doesn’t necessarily mean the robot successfully completed what it was supposed to do. So you want to ask both… “Is this trajectory clean?” and “Did the robot actually accomplish the task?” That distinction matters when you are building datasets for Physical AI. The interesting thing about Axis is that you can start seeing how the different repositories fit together. Continuum Task Gen helps create tasks. Humans can demonstrate those tasks through teleoperation. Those interactions produce trajectories. Then tools like Continuum Data Cleaning can help process and evaluate that trajectory data. Only after those steps do you really start getting closer to training-ready data. That’s a much more complete way of looking at robotics. It’s not just… collect as much data as possible. It’s… →collect → inspect → clean → validate → improve → train. Because if robots are going to learn from human demonstrations, the quality of those demonstrations matters. And that’s why I think the less glamorous parts of the pipeline, data cleaning, validation and evaluation, could end up being just as important as the flashy AI models everyone talks about. Tomorrow, I’ll take a closer look at what actually happens when you clean and process a robot trajectory with @continuumlabs_