Will robots explode in capabilities like LLMs did?
Dr Waku16 June 2026Watch on YouTube
Description
I interviewed David Watkins, who works at the AI Institute in Boston. This is a research-focused, privately held company which is experimenting with many types of robot platforms, especially Spots and robotic arms. Although many companies are creating robots, there are no standards or well accepted paradigms yet; the field is still in an exploratory mode. Robotic systems do not always copy biological designs, for example it would be hard to create a human style hand. This only adds to the data collection problem however. Robots do not work nearly as well as they could right now. Perhaps 10% of the capabilities of hardware is actually used by the software algorithms driving them. The biggest obstacle to making robots work is large-scale data collection. There is no way to gather ground truth data because we don't know the proper way for robotic systems to move in the world yet. For large language models, there was existing ground truth in the form of human written text. Hence, imitation learning was possible instead of straight reinforcement learning. But robotics systems have to learn from the ground up with RL. #robotics #machinelearning #ai The AI Institute https://theaiinstitute.com/ Universal Manipulation Interface: In-The-Wild Robot Teaching Without In-The-Wild Robots [paper] https://arxiv.org/abs/2402.10329 Sensei: Robotic Training Data at Scale https://www.ycombinator.com/companies/sensei Chelsea Finn’s robot easily sautes shrimp https://ee.stanford.edu/chelsea-finns-robot-easily-sautes-shrimp Mobile ALOHA: Learning Bimanual Mobile Manipulation with Low-Cost Whole-Body Teleoperation https://mobile-aloha.github.io/ Data Scaling Laws in Imitation Learning for Robotic Manipulation [paper] https://arxiv.org/abs/2410.18647 RoboCrowd: Scaling Robot Data Collection through Crowdsourcing [paper] https://arxiv.org/abs/2411.01915 Ego4D: Around the World in 3,000 Hours of Egocentric Video https://www.cs.utexas.edu/~grauman/papers/ego4d-cvpr2022.pdf David Watkins [LinkedIn] https://www.linkedin.com/in/david-watkins David Watkins [Google Scholar] https://scholar.google.com/citations?user=2S7jVrMAAAAJ 0:00 Intro 0:31 Contents 0:44 Part 1: The AI Institute and background info 0:52 Company founded by Marc Raibert 1:30 Relation to Boston Dynamics 2:16 Deep pockets and 100% research 2:51 Comparison to university environment 3:50 The effect of scaling models to internet scale 4:21 Will there be a ChatGPT moment for robotics? 5:11 How researchers can get involved 5:45 Part 2: Robots, including Spots and arms 6:49 My visit to the AI Institute 7:16 The Spot robot and its capabilities 7:43 Full low level control for the Spot robot 8:27 Different types of arms and impedance-based control 9:48 Arm detecting if it's bumping into something 10:51 PhD thesis: moving around and taking pictures 11:54 Perceptual cameras are used a lot for robot arms 13:05 Having too many cameras weighs too much 13:50 Robot sensors might go on the head 14:54 Can we just use all this camera data for foundation models? 16:08 Sampling data techniques were very important for ChatGPT 17:14 What is the structure of robotics data we're learning? 18:29 Is robotics hardware solved? 19:34 Are there concerns about robot safety right now? 20:37 LLMs don't experience the world 21:23 Does an AGI system have to be embodied? 22:15 AI safety through access to embodied data? 22:59 Part 3: Large scale data collection 23:08 How can you scale up robotics data collection? (technical) 24:10 Translation of big words 24:33 Sensei Technologies: Mechanical Turk for robotics 24:57 Chelsea Finn at Stanford, robotic tele-operation 26:10 Data scaling laws for manipulation 26:41 Boring tasks like picking up cups of water 27:34 You want as many environments as possible 28:06 The Google arm farm wasn't actually that useful 29:01 Analogy: person playing only one video game 29:46 Can we use video streaming etc as data sources? 30:47 Comparison with RLHF in ChatGPT 31:38 RLHF data once collected continues to exist 32:45 Number of dimensions in robotic control 33:50 Density of possible answers vs supports interpolation 34:25 State space explosion, computationally intractable 35:08 Large language models use higher dimensional structure 35:51 Handling soft exceptions 36:34 How large are the neutral networks? 37:01 Reinforcement learning needs small models 37:45 Large language models have existing structure when applying RL 38:51 Robots have no ground truth data 39:21 Imitation learning vs reinforcement learning 40:38 Is this the secret behind robotics progress? 41:19 Prosthetic hands just like human hands? 41:55 Hands have to be mechanically feasible 42:54 Robotics still in an exploratory phase 43:51 Active learning research in robotics 44:49 Don't even have a common robot vocabulary 45:00 Can we get robots to learn like humans do? 46:06 Advice for people in general 46:58 More accessible starting point for robotics 47:44 AI Institute is hiring interns etc 48:13 Outro
What you'll learn
- Robotics is still in an experimental phase without standardized paradigms, unlike the mature LLM landscape.
- The biggest bottleneck in robotics is scaling data collection, because there is no existing ground truth data like human-written text for LLMs.
- Current robots use only about 10% of their hardware capabilities due to limitations in software algorithms.
- Imitation learning was possible for LLMs thanks to existing human-written text, but robots must learn from scratch using reinforcement learning.
- The robotics community still lacks standardized vocabulary and common approaches, which hampers further progress.