15 claims in “robotics”
Unitree's robot dog starts at about $1,500, while the comparable Boston Dynamics robot dog costs $50,000 to $100,000, and the Unitree dog can do backflips, climb stairs, and run a built-in LLM.
Robotics transformers such as Google's RT-1 benefit from pretraining on huge text corpora to build world understanding, then adding robotic sensor inputs as another modality, which transformers handle naturally.
Robotics AI lags because robots cannot afford to fail a million times to learn tasks like walking, and simulations fail to transfer well to the real world, so training data is scarce compared with image datasets.
Robots need very fast response times because a one-second round trip to a server is too slow for physical control, so latency is critical, and this is why transformer robotics has not reached the mainstream despite Tesla using vision transformers.
Guo says the central robotics problem is data: fully general robots need data reflecting real-world environments and tasks, and the key innovation is collecting it cheaply and transferring that cheap data into model learning.
Sunday Robotics' team believes general semi-humanoid robots will be doing tasks in people's homes in beta by the end of this year, and Guo says that belief is one she and most of the field did not hold before, with robotics insiders now seeing it as a question of when, not if.
Two very young Stanford PhD founders, Tony Zhao and Cheng Chi of Sunday Robotics, have contributed most of the interesting ideas in robotics AI over the last four years by treating robot data collection as a cheap, practical technical problem.
Guo says Sunday Robotics' founders were impressive because in under two years they went from a basement prototype to a full-stack system through hundreds of hardware iterations, arguing that nothing is true until it is shipped.
Winfree and Lulu Qian built a single-molecule DNA robot whose domains walk on a surface via programmable Watson-Crick interactions, pick up cargo, and drop it at a designated location, with all control encoded in DNA informational barcodes.
A person can learn to teleoperate any humanoid or robot arm within hours, and robotics has not become a multi-trillion dollar industry because AIs learn far less efficiently than humans, so even millions of hours of demonstrations are insufficient for complex open-ended tasks.
Physical robot dexterity such as swishing a wine glass remains nowhere close for models, but robots do not need that level of skill to be valuable, since picking up and placing known objects can generate data and create value.
Naively showing a multimodal LLM a picture and asking what it would do fails, because the model does not know it is a robot, what the robot looks like, or what is going on, so connecting LLM knowledge to robot control is a real technical challenge.
Progress in form factor innovation has been constrained in traditional robotics because building a new robot requires characterizing dynamics and system identification; a general robotic foundation model would let anyone assemble a robot in a garage and get it moving, which could create a powerful engine of experimentation.
The Roomba is the best-selling consumer robot of all time, which is surprising.
Humanoid robots remain mechanically challenging beyond AI - the engineering of the physical body is the primary bottleneck, not the intelligence system