Robot Dogs Can Now Grab Things Even When They Can't See
Machines that fetch and carry are getting closer — even when cameras are blocked or blurry.
A team of researchers has taught a robot dog with a robotic arm to grab and carry objects even when its eyes — its cameras — are giving it bad information. Think of it like reaching for your coffee cup while wearing fogged-up glasses in a dim room. Most robots today only work reliably when their cameras are clean and their view is clear, which is why they struggle in messy real places like warehouses, farms, or cluttered homes. This team focused on exactly those bad conditions.
They built two things. The first is a set of tests, called DeViGrasp-Bench, that deliberately mess up the robot's vision: things blocking the camera, missing chunks of the image, noisy depth readings, and objects that seem to jitter around. It also throws in unfamiliar objects and rough ground. The second is the robot's brain, DeViGrasp-Net, which works a bit like a student learning from an expert. The expert sees everything perfectly during training, while the actual robot learns to cope with blurry, incomplete pictures — and to remember where an object was a moment ago, so it can recover if it briefly loses track.
The results are promising but not perfect. On the hardest test setting, the robot succeeded 62.3% of the time, beating two comparable systems by 16.1 and 4.3 percentage points. On the very hardest setting, its lead over the closest competitor grew to 10.5 points. In other words, it is clearly better than what came before — but it still drops the ball roughly four times out of ten when conditions get rough.
Why does this matter outside the lab? Robot dogs with arms are being pitched for warehouse picking, construction inspection, farming, and eventually household chores. Their biggest weakness has never been walking — it's understanding what they're looking at. A robot that keeps working when a lens gets dusty or a box blocks its view needs far less human babysitting, which translates into lower costs and fewer stalled operations. This research is a solid step in that direction, not the finish line.
- A four-legged robot with an arm learned to pick up objects even when its cameras are blocked, blurry, or noisy — conditions that normally make robots freeze.
- In the toughest test setting it succeeded 62.3% of the time, beating two rival systems by up to 16 percentage points.
- It still fails about four times in ten under hard conditions, so robots doing your chores unsupervised are still years away.
Why It Matters
Robots that keep working when cameras get dirty or blocked could soon take over repetitive lifting and fetching jobs.