I've been thinking a lot about making robot learning scalable. To me, that means learning on the job, and using the human help still needed efficiently. In Mulligan ⛳️, we focus each data-collection round on the initial states where the robot fails. This happens in deployment, with the robot doing the task and a person stepping in when needed. On 3 real tasks and 2,550 blind evals, it improves on uniform collection at the same budget.

In language models, turning a generative model into a critic is simple: pass in the output and prompt for a score. Robot policies can't do this - they output physical actions & there's no built-in vocabulary to predict scores. So when it comes to RL on robot policies, we typically train a separate critic from sparse rewards - blind to everything the policy learned from the actions in the pretraining data. We think this has it backwards - can a critic learn from those actions the way the policy does? (1/n)

I finally baked a long-form audit of “robot scaling laws.” AI agents did a lot of the digging; I did a lot of checking, rewriting, and arguing with the drafts. I reviewed 19 public results related to robot scaling. The short version: scaling curve ✅ scaling law ❌ Here is the evidence funnel: 19 public results reviewed → 16 report multi-scale curves → 13 keep the main variable and evaluation reasonably controlled → 1 tests a held-out loss forecast → 0 completes the full loop from fitting loss, to forecasting the next scale, to validating that forecast on real robots from the same model family. This distinction matters. In LLM scaling experiments, data volume D is usually measured in tokens. Token count does not tell us how much new information the data contains, but it remains a relatively stable training unit when the tokenizer, data pipeline, and distribution are held fixed. Robotics does not yet have an equivalent common unit. Different studies use recorded hours, trajectories, frames, action steps, or action tokens. Even the same “one hour” can produce very different training examples under different sampling rates, action representations, chunking schemes, filtering rules, tasks, environments, and embodiments. The reported objectives and y-axes vary even further: held-out denoising or vector-field MSE, decoded action error, thresholded accuracy, simulated progress, task completion, cycle time, and real-robot success all answer different questions. A few examples: — Figure Helix 2.5 provides the only held-out loss forecast I found: the first three Index-data scales are used to predict offline action loss at 8×. But its 9%→56% real-robot result comes from a separate with-vs-without-pretraining experiment, not the endpoints of that loss curve. — Dyna-2 scales egocentric human video from 1k to 1M hours. Offline acc@0.5 moves from roughly 40% to 47%, while post-trained robot performance rises from 20% to 53% on its normalized metric. Scaling helps, but the public results cannot tell us whether the final 900k hours were worth the cost. — Xiaomi-Robotics-1 reports that, with a fixed 5B model, increasing UMI data from roughly 2.5k to 20k hours raises mean success from 53% to 75%. At fixed data, moving from 5B to 10B adds only four points, from 75% to 79%. Useful evidence, but the full model × data grid is still missing. — Sunday ACT-2 shows that a quality-selected 12.5% subset can outperform a uniformly sampled 50% subset. Nominal scale is not enough: data distribution, repetition, and coverage can matter more than another hour of collection. My current view: A general law spanning tasks, environments, and embodiments will probably arrive later, because robotics still lacks a shared training and validation distribution. But deployment may not need to wait for a predictive scaling law. If a system is reliable enough on a concrete task at an acceptable cost, products may arrive before a general law does. Perhaps robotics is not waiting for a GPT moment, but an iPhone moment: hardware, software, data, and intelligence finally coming together in a product that works reliably in the real world. Full visual audit, original figures, evidence boundaries, and sources: minghuanliu.com/robot_scalin…

What if the jagged frontier is mainly math + code (which you can push arbitrarily far with RLVR), and everything else starts to plateau because it is still bottlenecked by human generated data? Model performance in non-verifiable areas has kept improving steadily, albeit much slower than for math and code. But is that steady improvement a side effect of a higher G (itself driven by RLVR), or only a function of the amount of new human data getting injected into training (which is still continually happening on a massive scale)? A lot of things depend on the answer to this question

finally the llm people are catching on to what we robot people have known for years jk, really cool work from the great @wen_kaiyue and @LukeBailey181 💅

I have a Continuous Learning benchmark where models attempt to learn to play chess. They are given a /goal of learning and improving playing against a Stockfish opponent in 200 games. They can choose the difficulty, take notes, whatever they like - except cheating (e.g. using a chess engine) of course. So far the improvement in Elo has been negative for Astra. Tiny bit positive for Opus, but could also be random. I've started Astra off sooner, so it finished its 200 games already, Opus is still playing. Site here to watch how they are doing: ai-learning-to-play-chess.su… The reason why this is interesting is that while we obviously don't have continuous learning, at the back of my mind I was thinking that maybe models can simulate it through self-scaffolding. Turns out not so much at least in this context. Perhaps it's a solvable problem and we don't need 'true' self-learning for models to learn in some way. ------- Just a note, the idea for the benchmarks belongs to someone else, but I don't want to use their name to give this more weight without permission.

AI can increasingly make progress on open problems like Navier-Stokes. Yet defining a new problem still relies on human research taste. Introducing ScholarCatalyst, a benchmark built from AI researchers’ firsthand accounts of what inspired their work.

feel free to help me get my clanker cleaner faster by using my referral link: tau-robotics.com/waitlist?re…

one nice things about scaling laws is that at least we know what won't work

woke up today with the unmistakable feeling that luke's work so nice I gotta retweet twice

New research: a hand is now a standalone robot! I love this project involving @Ske_Vinci, @HaochenShi74, @SongShuran, @dingmyu, and others. Called Handroid, it is an open-source, desktop-scale research robot (0.33 m tall, 2.05 kg, 27 DoF) that is both a dexterous anthropomorphic hand and a walking mini-humanoid. How cool is that ?! The same articulated electromechanical modules reconfigured allow the fingers to literally double as limbs. It is able to perform dexterous grasping, in-hand cube reorientation, pick-and-place, pouring, and bipedal squat/walk/turn, plus long-horizon tasks where it reconfigures its own embodiment (between bipedal humanoid and hand) mid-run (dock, locomote, manipulate). The project publishes GitHub code, OnShape CAD, and a Google-Sheets bill of materials, fully open, 3D printable -> you can build your own todat I really like that it is the most genuinely novel mechanical idea I have seen in a long time: not two robots in a box, but a robot whose fingers become its legs. Worth mentioning: Shuran Song's lab (@SongShuran) created UMI, the cheap hand gripper that became a standard for data acquisition. -> the people who built the dominant data-collection interface are now building cheap reconfigurable hardware. I believe this is a tell about where the frontier academic labs are now pushing the frontier: not another VLA, but better/cheaper embodiments to study on. Handroid's ability to transform between hand-mode and humanoid-mode is a genuinely different capability from every other robot we have seen lately, as most other robots have fixed embodiement, besides the ability to pick and use a tool. Morphology as a controllable variable is genuinely new, afaik. Enjoy watching this little robot as much as I do. Here it is able to plan a long horizon task, decompose it in steps, and behaves sequentially as either a hand or a humanoid (4x sped up):