The Data Problem in Robot Learning Is Getting Solved, Just Not the Way You'd Expect
After years of watching the industry chase bigger datasets, researchers are finally getting clever about making smaller ones work harder.
Crédito da imagem: Lottie animation by Centre Robotics (LottieFiles Free, used with credit). · source
Look, here's the thing about imitation learning for robots: we've been doing it wrong for years. Not completely wrong, mind you, but wrong enough that I've watched a lot of promising projects die on the vine because nobody could afford to collect fifty thousand demonstrations of a robot picking up a coffee mug.
I've been following a cluster of recent papers that suggest the field is finally wising up. And honestly, it's about time.
The Instrumentation Approach Actually Makes Sense
When I was at Kuka, we spent ungodly amounts of time trying to figure out why a perfectly good vision system would fail on tasks that seemed trivial. Turns out, cameras don't tell you everything. A team from what looks like a European research group just published work on instrumented imitation learning that basically puts sensors in the objects themselves, not just on the robot.
They tested it on clothes hanger insertion (not the most glamorous task, I'll admit) and found that policies with access to instrumentation data outperformed vision-only approaches by 14 to 25 percentage points. That's not a rounding error. That's the difference between a system that works and one that gets shelved.
What's clever here is the second part: they used the instrumented expert to generate additional training data, then trained a vision-only student policy on that enhanced dataset. The student ended up matching the instrumented expert's performance. So you get the benefit of all that sensor data during training, but you don't need to ship products with instrumented objects. I called my old colleague at Siemens about this approach and he was, well, cautiously interested. Which for him means excited.
Cobertura relacionada
More in Industrial
New research tackles the uncertainty problem in monocular depth sensing, and after 12 years of watching vision systems fail in warehouses, I have thoughts.
Robert "Bob" Macintosh · 26 mins ago · 3 min
While everyone's chasing humanoids, researchers just solved problems that have plagued factory robots for decades.
Robert "Bob" Macintosh · 26 mins ago · 4 min
A batch of new research on robot learning from demonstrations looks impressive on paper, but I've got some questions about what happens when these systems hit a real factory floor.
Robert "Bob" Macintosh · 27 mins ago · 4 min
Two new papers tackle the same old problem I've been watching for decades, and I'll be honest, one of them actually impressed me.