01 / The frontier
The next decade of robotics will be defined by systems that handle dexterous tasks at human speed — working alongside people in real time, in environments that don't hold still. The hardest part isn't capability in a demo; it's reliability in the real world — closing the long gap between a system that works once and one you can depend on.
Avolla is building those foundations.
02 / The problem
The gap is not sim-to-real. It's demo-to-deployment.
- 01 Robots aren't reliable enough to deploy.
- 02 And today's recipe makes learning on the job slow and tedious. (long tail, teleop)
- 03 So there's no field data to improve them.
Three reasons the gap persists
- 01
One big model is the wrong shape
Slow to update, too big for the robot, too fragile.
- 02
The behavior cloning ceiling
Bounded by the demonstrator; errors compound and the long tail never closes.
- 03
Contact-rich work, controlled by vision alone
The task is defined by force and touch; the policy never sees them.
03 / Our bets
Human references, simulation, and real-world adaptation, in one loop.
- 01
An agentic loop of skills at the edge, improving in the field
Each new task strengthens the skills; the effort to learn the next one falls.
- 02
Learn from human references and simulation
People already perform the tasks. Simulation provides coverage and robustness.
- 03
Lean into high-DOF hands and rich sensing
Dexterous hands with skills that close the loop on more than pixels.
04 / The company
A frontier lab that ships.
We build the intelligence and put it on robots in the field. We drive deployments end to end, partnering where it makes them better. The robot in the field is both the model's data engine and the real test of whether the demo-to-deployment gap is closed.
05 / Founder
How my work shaped Avolla's mission
I spent the first half of my 15 years at Google shipping on-device ML to billions of Android devices. That taught me what reliability means at the edge: a model has to deliver on the device in someone's hand, within its time, memory, and compute.
I spent the next seven and a half years in robotics at Google Brain and DeepMind, where I led the table-tennis robot program. Beating a human player took fast, precise returns, rally after rally, adapting as the match unfolded. I look for the same in a manipulation skill: adjusting when an object moves unexpectedly and still finishing the job, time after time.
I also co-founded and co-led Open X-Embodiment, which explored how robots could learn from experience collected on other robots. Leading data quality for Gemini Robotics taught me how much depends on having the right data. Together they left me with two questions. When a robot fails, have we captured enough about the attempt to know what it needs to learn? And do we have a recipe that learns from it quickly?
My PhD in nonlinear control taught me to trust feedback over a perfect model. At Avolla, I'm bringing all of that to dexterous work: recognize a failure and recover, close the loop quickly so the robot learns to handle it reliably, and carry what it learned to the next job.
06 / Join
Join the founding team
We're looking for people who trust their own judgment over the consensus, who are willing to go against mainstream methods when the evidence says so, and who believe a small, focused team can make a huge difference.
Tell us what you've built: hello@avolla.ai