RobotyThe product
How Roboty works
Set a goal, try an approach, inspect what happened and decide what to change.
Start with a task you can understand
Can this arm reach that target? It’s a small question, but answering it takes more than an animation. The target we choose has to reach the controller, the simulation has to run, and the result has to tell us whether the arm actually met the goal.
That’s where our current development loop starts: supported robot-arm simulations on a Mac, with an iPhone interface for authoring and reviewing experiments. Roboty is still in development.
Read about the first reach experimentKeep the question connected to the result
Choose a target
Set up an experiment and save the goal you want to test.
Try an approach
Run a supported simulation against that saved goal.
Inspect and change
Review the outcome, keep the history and change something meaningful for the next attempt.
A failed attempt can be useful if it helps us understand what to try next. Keeping the setup and result together lets us revisit the work without relying on a success message or our memory of what happened.
How we describe the evidenceUse an app or bring your own agent
The iPhone is our first native interface. We’re also building access for external AI agents, so people can use their own assistant to work with supported tasks and inspect the same results.
Selected local agent-access routes have already run authored simulation tasks in development tests. Compatibility with particular assistants and the experience of using them still need more testing. We’re working toward a choice of interfaces around one experiment history.
From simulation to physical work
Simulation gives us a way to explore a task before working with a real robot. Physical operation brings different questions about calibration, measurement and stopping the robot. A simulation result doesn’t answer those questions.
Our physical starting point is the SO-101 arm. Its operation through Roboty remains to be validated on the actual setup. We’ll share those results as we have them, including what doesn’t work.
We reuse existing robotics tools where they help. The work is making the whole experience more capable and easier to understand, from setting up a task to figuring out why it failed.
Follow the work on the blog