Microsoft just announced it’s redirecting AI Frontiers toward scaling. Everyone is racing to build bigger datacenters. So let me tell you what you can do with a robot and a laptop.
One of my students claimed there was no meaningful research he could do without the cluster. Of course there is. Not just in mechanical design and fabrication - one of my students is iterating on a UMI hand for our Spot that it can grab in its gripper and use to transfer policies demonstrated on a UMI to our most mobile robot that is the easiest robot we have to take places. But also real research ideas that would advance the state of the art and make a robot do something it couldn’t do before. I’m a computer scientist, and so my zone of excellence is finding clever ways to improve the design and architectural decisions around robust programming. I want to find clever ways to utilize desktop sized compute to do more impressive things, which always have the potential to be scaled later.
Modern AI methods leverage data and compute to learn models for sensing and acting. I toured our machine room and the home of our cluster at Brown, and we saw row after row of racks of GPUs with a jet-engine level of noise from cooling systems. And our cluster is small, nowhere near as large as our peer institutions, much less what is available in industry. Even so it is in constant use as students train models and run experiments.
Probabilistic Programming. Long ago there was extensive research into probabilistic programming by Noah Goodman, Vikash Mansingha, and Dan Roy. This work has continued to be explored to this day, but has lacked a killer application. The idea of probabilistic programming is that you can define a probabilistic graphical model via code (e.g., in Church, it was Scheme code), and automatically derive an inference algorithm by leveraging automatic differentiation and MCMC or variational techniques. It resulted in some cool cognitive science results but never found a killer app in the real world. The potential for robotics is uncertainty tracking and combining information from disparate sources via the common currency of probability. Before we had AI models, writing the program itself was challenging: a person had to encode the problem model in a formal representation, and we know this is hard. But now, with AI coding, we can rapidly and automatically generate the program itself. And leveraging the conjugate prior (e.g., Dirichlet and its ilk) can provide leverage for calibrated probabilities that can be meaningfully combined for an overall estimate. So here is how to start: write FastSLAM in a probabilistic programming framework. Extend it to a POMDP that chooses actions based on a reward function for visiting every place in the environment. Benchmark the performance in a 2D gridworld and then try it on a real robot like Spot or a Duckietown car. Then apply it to the specific case of in-hand reconfiguration for the ORCA hand, where the tensions/position of the fingers is unknown due to cable stretch, and use a Gibbs sampling or MCMC approach to plan behaviors to reconfigure the cube. In this formulation physics is part of the model-based probabilistic program, it’s differentiable, and it also cleanly incorporates perceptual uncertainty. Writing this program would have been a lot of work before AI coding models but is approachable now. I predict that it will take more runtime compute than a learned policy, but be more robust to noise.
Differential state estimation and control. Classical robotics decomposes the problem into state estimation and control. The Kalman filter is a classic choice for state estimation because it is provably optimal when its assumptions are met. PID is the classic choice for a controller, not because it has proofs but because it works well in practice. Use AI to write a differential version of this stack (via a differentiable object detector and convolutional filter) and benchmark whether this classic structure outperforms diffusion policy on LfDs. I predict it will work better with less data but asymptote lower following the pattern from Michael Jordan and Andrew Ng’s paper, which we reviewed on this blog earlier in Where the Curves Cross.
Trash pickup. With new robotic capabilities come new opportunities for a wide array of verticals. A systems robotics challenge I’ve been thinking about is trash pickup. We now have robust models for grasping a wide variety of objects. And with litter pickup, you don’t need to worry about breaking the object, and if you miss it, well it’s not great, but it’s also not a total failure for the job. This space isn’t unexplored, but I believe it is underexplored. Research challenges include robustly detecting trash and sorting objects into landfill and recycling. Can you imitate half asleep Chris, the Youtubber who picked up a ton of trash, and build a robot that does it every day?
I’ve deliberately picked ideas that none of my students are working on. There are many more! The key is sitting next to the robot, making it do the thing, and explaining why. You don’t need a data center for that.
Hi Stefanie! I am working on trash pick up as a part of my PostDoc fellowship. We are partnered up with MediMRF an early stage startup that sorts trash in hospitals. As you mention, we work with very little compute.
Hi Stefanie! Love this article. I had very little compute and presented my research at CoRL this week. Agreed with all your points, especially around Sutton recently clarifying his point on the bitter lesson. There are a ton of problems in this space that more efficient algorithms and clever engineering can address, e.g. Supermimic was a great robotics paper of a very simple model (3 layer MLP world model).
Trash pickup is a fantastic soft body problem because if a failure occurs, it's not catastrophic. I also think it has great applications in both industry and nonprofit work, e.g. cleaning up polluted rivers and oceans (that video from Chris is very cool).
Dyna Robotics was a major hit at CoRL and I think one of the reasons they're successful is the problem they've chose to work on, i.e. towel/laundry folding. It makes closed loop evaluation easier than more critical use cases (e.g. industrial parts assembly, which really have to ensure the 9s of reliability because failures are costly).
As someone going back to grad school next year for CS/robotics, this article felt very aligned with the research I'm thinking of tackling next + my own approach to research.