Robot learning stopped arguing about policies and started arguing about data

The pipeline is the paper

CoRL 2026 accepted 687 papers from 2,094 submissions, a 32.8% acceptance rate, and the strongest robot learning papers at the conference focused on data pipelines rather than new policy architectures. That is the read from Avala Research, and it sets the tone for where robot learning is heading: the binding constraint has shifted from feasibility to scale and diversity, and scaling human-to-robot learning now relies on conversion stacks, pose estimation, retargeting and quality filtering rather than raw video. m:…8761

The numbers behind the shift are about labeled hours, not recorded hours. EgoScale pretrained a vision-language-action model on 20,854 hours of egocentric human video, and validation loss on human action prediction falls log-linearly as pretraining hours grow. Scaling laws for robot learning are increasingly tied to labeled hours rather than raw recorded hours, with volume helping to absorb label noise. m:…8761

Curation is changing shape too. It is moving from rule-of-thumb filters to measured scores, with techniques like influence functions and progress weighting enabling more efficient training. ATHENA reduced influence computation for a billion-parameter VLA by 313.4x, which is the kind of unglamorous infrastructure work that makes large-scale curation affordable. And WARP-RM improves task throughput by weighting action chunks based on task progress without human annotation. m:…8761

Time is a deadline, not a detail

The second theme from the conference is that the world does not pause while a robot thinks. Hang Zhao showed that direct frontier-model policies completed 15% of tasks when inference time counted toward a 50-second deadline, and agentic policies completed 20%. Four approaches cleared 60% of tasks when thinking time was free, which is the gap between a demo and a deployment. m:…1764

Zhao splits robot control into three layers: reasoning in seconds, movement coordination in tens of Hz, and contact and force control in hundreds of Hz. The layered answer is the point. A robot that reasons well but moves late still fails, which is why the field is building control stacks that keep the fast loops fast while the slow model thinks. m:…1764

Memory is the other half of the time problem. TEMPO increased bottle handover success from 38% to 74% by adding a video-derived motion summary and compressed movement history to a VLA model, and MEM enables robot policies to perform tasks lasting up to 15 minutes by combining short-term video memory with longer-term text summaries. RoboTTT stores interaction history in fast weights that update as the robot acts, and allows policies to use longer histories without raising per-step inference cost. m:…1764

Failures are training data now

Failed rollouts are becoming valuable training signals when accompanied by metadata on execution quality, mistakes, and recovery segments. A model from Physical Intelligence uses context conditioning to treat failed rollouts as training signal, and RoboTTT improves performance by 87% over single-step-context baselines by extending visuomotor context to 8,000 timesteps. The failures a robot used to hide are now the data it learns from. m:…8761

Synthetic data is pulling in the same direction. GRAIL generates humanoid loco-manipulation data without human video, achieving 84% real-world success on object pick-up. And T-Rex improves average success on manipulation tasks by over 30% by incorporating tactile signals into the action loop, a reminder that the missing sense in most robot data is touch, not sight. m:…8761

DexForesight makes the same point from the dexterity side. It improved the overall success rate of the base policy from 49.4% to 60.5% on 11 DexJoCo simulation tasks, and achieved a 58.3% success rate on three real-robot dexterous manipulation tasks, against 50.0% for the strongest baseline. The world model captures the motion of fingers occluded from cameras, and at deployment the policy operates on RGB observations, language, and robot state without depth input or explicit future prediction. m:…1806

Memory can lie about time

There is a catch in all of this, and it is about how memory is indexed. Time-based memory indexing fails to distinguish between waiting and working, and brief pauses or wrist fidgeting in RoboMME counting tasks cause early stopping in VLA policies. On one RoboMME task, success dropped from 47 to 16 of 50 due to pauses, because the policy counted an interval of inactivity as work done. The paper is a preprint, so the findings are the authors' own and not yet peer-reviewed. m:…1808

The fix is to index memory by events rather than by the clock. Indexing memory by events recovers half or more of the performance loss without retraining, and withholding pause frames from memory restores success rates. Removing time encoding does not cure the failure, and training with injected pauses does not remove it either, which suggests the problem is structural: elapsed time predicts progress in demonstrations because they proceed at a steady pace, and the real world does not. m:…1808

The newest direction skips weight updates entirely. World In-Context Learning is a paradigm where a frozen model rehearses tasks in a simulator and writes feedback into its context, and RoboScript is a modular infrastructure that implements it for agentic manipulation. RoboScript reaches 83.4% mean success across 50 tasks of RoboTwin 2.0 and exceeds the strongest trained action policy by over 20 points, without gradient updates. The robot learns in context, rehearses in simulation, and deploys what it learned as a Task Memory on new scenes. m:…2396