Doppler LiDAR and Vision Systems Are Finally Catching Up to What Factory Floors Needed a Decade Ago
Two new research papers show promising approaches to obstacle avoidance, and I'm cautiously optimistic we're getting somewhere useful.
By
When I was at Kuka, we used to joke that our robots could weld a perfect seam on a moving car body but would panic if you rolled a trash bin past them unexpectedly. That was 2014, maybe 2015. The state of dynamic obstacle avoidance was, let's say, not great. So when two papers crossed my desk this week showing genuinely clever approaches to the problem, I felt something I don't feel often anymore: actual optimism.
The first one comes from a team working on what they call DPNet, published on arXiv. The core idea is using Doppler LiDAR, which gives you not just distance measurements but instantaneous velocity of points in the environment. I'll be honest, when I first heard about Doppler LiDAR applications in robotics a few years back, I thought it was a solution looking for a problem. Regular LiDAR seemed good enough. But the DPNet folks have built something that actually makes use of that velocity data in a meaningful way.
They've created what they call a Doppler Kalman neural network (D-KalmanNet, if you want the full mouthful) that tracks obstacle states even when you can only partially observe them. This is the real world, after all. Things get occluded. Sensors miss readings. The system then feeds those predictions into a model predictive control framework that auto-tunes itself at runtime. Now, I've seen plenty of MPC implementations over the years, and the auto-tuning bit is what caught my attention. Most of the systems I worked with required careful parameter tuning for each deployment environment. If this actually works as advertised, it could save integrators a lot of headaches.
The second paper is about quadcopter navigation, which might seem unrelated to my usual beat, but bear with me. This team from (I assume) somewhere with access to a lot of outdoor testing space has built a vision-guided system that uses stereo depth cameras and visual-inertial odometry to navigate through obstacles without GPS or telemetry. The paper is on as well. What's interesting here isn't the drone application specifically, it's the approach: they trained the policy in simulation using reinforcement learning with a privileged learning setup, then achieved what they call zero-shot transfer to real hardware.
関連記事
More in Industrial
The Apple supplier priced its shares at the maximum and still had to turn away demand, which tells you something about where hardware money is flowing right now.
James Chen · 25 Jun · 5 min
Prime Day deals on Echos and Ring cameras are fine, but let's not confuse consumer gadgets with the serious robotics work happening in warehouses.
Robert "Bob" Macintosh · 25 Jun · 3 min
Amazon's CEO made his first India trip and left behind a $13 billion AI commitment and an aggressive quick-commerce expansion. The numbers are real. The execution is the hard part.
James Chen · 25 Jun · 6 min
A wave of arXiv preprints this week tackles one of manipulation's oldest problems: how do you get a robot to learn from imperfect, incomplete, or just plain missing data?
