Robots learn from watching people

Scale from human video

Open-HumanWAM is an open recipe for scaling human and robot experience for embodied foundation models, built on human video at scale. The preprint reports an empirical log-linear scaling law linking pretraining data hours to downstream robot success rates — more hours of human video, predictably better robots — and says hybrid human-robot pretraining unlocks broad zero-shot generalisation across simulation and physical benchmarks, achieving 84.60% success on RoboTwin 2.0 Clean2Random. The model HumanWAM-Goal exhibits emergent embodied in-context learning, dynamically adapting manipulation behaviour to unseen visual goals without parameter updates. m:…4280

The idea is simple and important: robot data is scarce and expensive, human video is abundant. If human experience transfers to robot policy at a predictable rate, the field can plan its data collection the way language modelling planned token counts. The log-linear law is the kind of result that turns a research direction into a budget line. Whether it holds across tasks beyond manipulation benchmarks is the question that will decide how much money follows it. m:…4280

Memory and space

HeLM, a hierarchical vision-language-action framework, attacks a different limitation: most manipulation models are Markovian, meaning they act on the current observation and forget what happened. HeLM maintains task-relevant history in explicit language through factorised event detection and event-triggered memory updates, and its high-level planner updates memory and generates action commands only when a task-relevant event is detected or the instruction changes. Evaluated across 10 non-Markovian real-world tasks with 56 variants, it improved intra-episode success by 12.8 percentage points over the strongest baseline and achieved 78% subtask success on compositional generalisation tasks. m:…6487

It is also cheaper to run. HeLM uses up to 4.5x less memory than the keyframe baseline at long horizons and runs 2.5x faster, and two-stage inference improved average intra-episode subtask success from 67.7% to 85.7% compared with a one-stage variant. Memory that is explicit and language-shaped turns out to be both more effective and more efficient than keeping every frame around. That is a useful combination for robots that carry their compute with them. m:…6487

SLA, the spatial-language-action model, adds the third ingredient: 3D awareness. It continually pre-trains and post-trains a pretrained vision-language-action policy, using spatial alignment to a frozen 3D perception model, feature-dependent gates and probabilistic state dropout to improve spatial understanding. It is initialised from a pretrained pi0.5 policy, and it preserves the original inference architecture of the pretrained policy. The point is that vision-language-action foundations often inherit visual representations from 2D vision-language pre-training, which is a poor fit for a robot that has to grasp things in three dimensions. m:…6484

Robots that write their own skills

PyRUA-REAL closes the loop. It is a framework for embodied recursive self-improvement that lets coding agents convert physical robot exploration into persistent, reusable skills: coding agents retain interaction experience as persistent memory and reusable executable skills, and programs generated by the framework can be deployed without an online agent for decision-making. It achieved an average success rate of 91.1% across eight real-world tasks and 84.4% across 42 tasks from the RoboDojo benchmark. m:…8296

The study also examines how task specifications and previously acquired skills influence robustness and task transfer, how execution feedback supports effective code refinement, and how agents collaborate during the improvement process. The underlying claim is that exploration should leave something behind — a skill, not just a memory of having tried. If that holds, each deployment makes the next one cheaper, which is the closest thing embodied AI has to a compounding advantage. m:…8296

What this adds up to

Four preprints, four different answers to the same question: where does robot experience come from? Human video at scale, explicit language memory, spatial awareness baked into the policy, and self-improvement that writes reusable skills. None of these is a robot you can buy. Together they describe a field that is converging on a recipe: pretrain on human experience, give the robot a memory and a sense of space, and let it turn its own exploration into code it can reuse. m:…4280 m:…6487 m:…6484 m:…8296

For builders, the practical signal is that the bottleneck is shifting from the model to the data pipeline. If Open-HumanWAM's scaling law holds, the teams with the best human-video pipelines win. If HeLM's memory and SLA's spatial alignment hold up, the architecture choices are settling. And if PyRUA-REAL's skill persistence works outside the lab, deployment becomes a flywheel rather than a cost centre. All four claims are preprints, so treat the numbers as early. But the direction of travel is consistent across all of them. m:…4280 m:…6487 m:…6484 m:…8296