The Conference on Robot Learning (CoRL) 2026 closes in Austin this week with 687 accepted papers from 2,094 active submissions, a 32.8% acceptance rate, with 21 orals and 72 spotlights. The program chairs put submissions at nearly three times last year's. Working through the accepted list left us with one conclusion: the strongest robot learning papers of the year are data pipelines. Their central contribution is not a new policy architecture. It is a method for turning human video into robot actions, for deciding which demonstrations stay in the training set, or for making a failed rollout useful.
The human-video results make the case most plainly. Three independent groups in the program pretrained on human-derived data at five-figure-hour scale or its equivalent: 20,854 hours, 18,561 hours, and about 630,000 clips. None of them trained on raw video. Each built a conversion stack first (hand and camera pose estimation, retargeting to a robot hand, rendering a robot into the frame, and quality filtering), and each reports gains on real robots from the converted data. One of the conference workshops states the premise outright: in human-to-robot learning, "the binding constraint has now shifted from feasibility to scale and diversity."
A disclosure before the guide. That reading favors the work we do: Avala is Fleet Learning Infrastructure for Physical AI, and conversion, labeling, and curation are a large part of it. So we have marked where the papers cut against it. One section below is given over to results that reduce dependence on newly collected human demonstrations, and two of the curation papers remove human annotation instead of adding it. Here is our opinionated guide to the papers we learned the most from, organized around what each one asks of the data underneath.
Human video crossed 20,000 hours, and labels carried it
Demonstration data is where our own work starts, so we begin with the papers that speak to it most directly.
EgoScale (oral; NVIDIA, UC Berkeley, and the University of Maryland) pretrains a vision-language-action (VLA) model, a policy that maps camera observations and a language instruction to robot actions, on 20,854 hours of egocentric human video. The authors put that at more than 20 times the scale of prior efforts. The word that matters in the abstract is "action-labeled." Every frame carries relative wrist motion and hand joint actions retargeted to a 22-degree-of-freedom (DoF) robot hand, recovered by running off-the-shelf SLAM (simultaneous localization and mapping) and hand-pose estimation over the footage. The in-the-wild portion spans 9,869 scenes, 6,015 tasks, and 43,237 objects.
Two results make the paper more than a large dataset. First, validation loss on human action prediction falls log-linearly as pretraining hours grow, with a reported fit of R² = 0.9983, and the authors report that this loss correlates with real-robot performance. Human data becomes an input a team can forecast against a budget, at least inside the measured range: the authors say they do not extrapolate beyond it. Second, the bridge to the robot is small. A mid-training stage on aligned human and robot data uses about 50 hours of human data and 4 hours of robot data, and the final policy improves average success by 54% over a no-pretraining baseline on the 22-DoF hand, with transfer to hands with fewer degrees of freedom.
Two other groups cross the gap between human and robot bodies in pixels as well as actions. Ego2Robot (Renmin University of China, ShanghaiTech University, and collaborators) retargets human actions, renders a robot arm into the egocentric frames, and applies multi-level quality curation, producing 18,561 hours of robot training data across 15 robot morphologies. Joint pretraining on the synthesized data and real robot data improves out-of-distribution generalization on an extended RoboTwin 2.0 benchmark, with the benefit confirmed on a real robot. HuRo (RLWRLD and Yonsei University) converts five human-video sources into about 630,000 robot-form clips (the paper's unit is the episode) and 142 million frames. Across four real-world manipulation tasks, pretraining on more of that data raised overall completion from 51.5% to 80.3% and out-of-distribution completion from 34.9% to 72.2%. The ablations are the part we would point a data buyer to: visual robotization helps out-of-distribution robustness, and pretraining with retargeted actions beats visual-only transfer.
Do as I Do (UC Berkeley) attacks the step all three depend on: reconstructing hand-object interaction from ordinary monocular video, first-person or third-person, and retargeting it into action sequences a multi-fingered robot hand can execute. The authors report better hand-object estimates than prior methods on datasets with ground truth, and they close with what they call "an efficacy playbook for practitioners collecting human data for manipulation." The abstract carries no headline number, so read the experiments before quoting one.
What this establishes, and what it does not. Three groups made different choices about how to cross from human to robot, and all three needed the same machinery to do it: pose estimation, retargeting, rendering or alignment, and filtering. That makes the labeled hour, not the recorded hour, the unit of scale. It matches the argument in what robot data costs, where we called cost per hour the wrong metric, and it extends the human-to-robot ratio we examined in from imitation to imagination. What the papers do not establish is how good the labels must be. EgoScale's authors describe their own pose estimates as noisy, and the scaling law held anyway, which is evidence that volume absorbs some label error. No paper here holds hours fixed and varies label accuracy. Until someone runs that experiment, "better labels beat more hours" is our inference, not a finding.
Curation got a measuring instrument
Curation is moving from rule-of-thumb filters to measured scores, and this year the measurement reached foundation-model scale.
At CoRL 2025, CUPID used influence functions (a technique that estimates how much each training example contributed to a model's behavior) to rank individual demonstrations by their effect on a policy's expected return. A diffusion policy trained on less than 33% of the curated data reached top results on the simulated RoboMimic benchmark. The open question was whether the method could survive billion-parameter models training on many tasks at once.
ATHENA (Shanghai Innovation Institute, Tongji University, Shanghai Jiao Tong University, and collaborators) answers it. Structured approximations cut influence computation for a billion-parameter VLA from 8,054.6 to 25.7 GPU-hours, about a 313.4× speedup, and added terms keep curation balanced across 50 jointly trained tasks so that no task is starved of data. ATHENA matches or exceeds full-data joint fine-tuning with 50% of the demonstrations in simulation and 66.7% of the data across six real-robot tasks. The scope matters: those fine-tuning sets are 9.34 hours in simulation and 6.90 hours on real robots. Nothing here shows influence scoring at the 20,000-hour pretraining scale, and an influence score is relative to one model and one evaluation. A demonstration that hurts this policy may help the next.
WARP-RM (UC Berkeley) curates inside the trajectory instead of across trajectories. A self-supervised model learns per-frame task progress from raw demonstrations, with no human annotation, and training then weights action chunks by how fast they advance the task. Discarding a whole demonstration because it contains a mistake also discards the recovery. This approach keeps the recovery and down-weights the stall. The authors report T-shirt folding throughput rising from 31.6 to 56.3 shirts per hour and bottle placement from 147.8 to 237.8 bottles per hour.
Scale Up Strategically moves curation upstream, into the collection plan. The authors measure which parts of an instruction a fine-tuned policy leans on as a shortcut. Across six foundation policies the ordering is broadly consistent: color dominates, then object, then spatial terms, with verb and size the least grounded. Shifting a fixed collection budget toward the under-grounded factors beats the baselines in simulation and on a real robot with half the demonstrations.
The same idea shows up where no human is involved. A Balanced Data Diet finds that in reinforcement learning (RL) with over one million parallel simulated environments, sampling starting configurations uniformly wastes a growing share of experience on states the policy has already mastered or cannot yet attempt. Concentrating on the frontier of current ability solves multi-terrain quadruped locomotion and contact-rich assembly that prior methods fail on. And MolmoAct2 (Allen Institute for AI, University of Washington, and collaborators) shows curation becoming part of publishing: its open release includes 720 hours of teleoperated bimanual trajectories alongside quality-filtered subsets of DROID and of SO-100/101 data.
Supports the thesis, with two cuts against the human share of it. Every method here treats "all available data" as the wrong training set, and two of them get equal or better policies from half to two-thirds of it. But WARP-RM and A Balanced Data Diet curate with no human labels at all. What all of them require is that each demonstration be individually addressable: a stable identity, a link to the recording it came from, and a join to the evaluation that scored it. That is the case for verified dataset releases, with one caution. Lineage makes attribution possible. It does not perform it; the influence computation does.
Failures became training signal, once they are labeled
π0.7 from Physical Intelligence is the paper we would hand to anyone who deletes failed rollouts. Its central idea is context conditioning: the prompt describes how a task was performed, not only what the task is. Alongside subtask language and subgoal images, a training example can carry metadata on speed, on execution quality (a score from 1 to 5), on mistakes, and on control mode. With that context, one recipe trains on demonstrations, autonomous rollouts that include failures, human interventions, egocentric human video, and web data. The authors' own framing is that diverse trajectories must be labeled with how the task was performed so that the model can put them in context. The capabilities reported in the abstract are qualitative: folding laundry on a robot that never saw the task, at a level the authors compare to expert teleoperators on their first attempts, and operating an espresso machine at a level that matches specialized RL-finetuned models.
The consequence for data operations is direct. A failed sequence with a recorded outcome, a quality grade, and a marked mistake is training signal. The same sequence without them is noise, or an example the model learns to imitate. Outcome labels, quality and speed grades, subtask boundaries, and mistake and recovery segments are fields we record on every delivered sequence, which is one reason we read this paper closely. The paper does not tell us what label error the method tolerates, and we would not assume the answer.
Memory Anchors (Stanford and Toyota Research Institute) makes a related point about old data. When a deployed policy learns a new task, a small subset of past experience does most of the work of preventing forgetting: the cases where a new task looks like an old one but demands a conflicting action. Excluding only the 10% of past data identified as anchors, before the replay buffer is sampled, increases catastrophic forgetting more than 4.5× on the LIBERO benchmark suites. Enriching the buffer with anchors cuts forgetting on the most conflicting task pairs by 63%. Which old sequences a fleet keeps is a curation decision with a measurable price.
MEM, the memory architecture we covered in four bottlenecks, three workarounds, appears in the program as a spotlight, and its long-term memory trains on summaries derived from subtask and outcome labels. The pattern across all three papers is the same.
Supports the thesis, on the condition in the heading. Failure data, intervention data, and old data all gained value this year, and in each paper the value is conditional on a record of what happened in the sequence. None of the papers measures the cost of producing that record, which is the number a program manager needs.
Minutes and context joined hours as scaling axes
DreamZero (oral; NVIDIA) is a world action model (WAM): built on a pretrained video diffusion backbone, it predicts future video and actions jointly instead of mapping observations straight to actions. On real robots the authors report over 2× the generalization of leading VLAs to new tasks and environments, from a 14B-parameter model running closed-loop at 7 Hz. Two transfer results matter most for data planning. Video-only demonstrations from other robots or from humans, 10 to 20 minutes of them, yield a relative improvement of over 42% on unseen tasks. And 30 minutes of play data adapts the model to a new robot body. We covered the debate over how much video generation a WAM needs in that same August post; two of the papers on the other side of it, ImageWAM and LaWAM, are also in this year's program.
RoboTTT (oral; NVIDIA, Stanford, and collaborators) adds context length as a scaling axis. Test-time training compresses history into fast weights, extending visuomotor context to 8,000 timesteps, which the authors put at three orders of magnitude beyond current policies, without added inference latency. On real-robot tasks it improves overall performance by 87% over a single-step-context baseline and is the only model tested to fully complete a five-minute, 10-stage assembly. The same model pretrained with 8,000-timestep context beats its 1,000-timestep version by 62%.
Put the year's data quantities on one scale and a ladder appears: 20,854 hours of converted human video for pretraining, about 54 hours of aligned human and robot data for mid-training, 6.9 hours for a real-robot fine-tuning set, 30 minutes to adapt to a new body, and 10 to 20 minutes of video to transfer a skill. These figures come from different papers and different systems, so read the shape, not the ratios.
These quantities span several orders of magnitude. All figures are author-reported and come from separate systems, so the chart shows a pattern across papers, not one pipeline.
Bar lengths use log10(hours) on a common axis from 0.1 to 100,000 hours. The final bar uses 15 minutes, the midpoint of the reported 10 to 20 minute range.
The planning implication is that hours stopped being the only axis. A team now decides how much converted human video to pretrain on, how much aligned data to bridge with, how much robot data to fine-tune on, and how long each recorded session runs. RoboTTT puts a price on cutting sessions into short clips: the context a policy could have learned from is gone. Our inference from DreamZero, not its claim, is that when minutes carry this much leverage, capture quality per minute (synchronization, calibration, and coverage of the target motion) matters more than it did when the unit was the thousand-hour batch.
Mixed for our thesis. These results shrink the amount of robot-specific data a new task or body needs, which cuts against any argument built on volume. They raise the stakes on which minutes are captured and how sessions are recorded, which is an argument for care over scale. Both readings are consistent with the numbers.
Where the papers cut against us: less new demonstration collection
The simulation-first and synthetic-first results get their own section, because they challenge a reading of the year built around collecting more human demonstrations. They do not all eliminate human-derived data: SuperMimic tracks a human-motion library.
GRAIL (oral, and a Best Paper Award finalist according to NVIDIA's project page; NVIDIA and UCLA) generates humanoid loco-manipulation data without a camera ever pointing at a person. It starts from fully specified 3D configurations (object geometry, camera, scale, environment depth, and a robot-proportioned character), uses video world models to generate human-object interaction, recovers metric 4D trajectories, and retargets them to a humanoid. The pipeline is virtual until deployment and produces over 20,000 sequences spanning pick-up, whole-body manipulation, sitting, and terrain traversal. Policies trained on that data and deployed on a Unitree G1 reach 84% real-world success on object pick-up and 90% on stair climbing.
SuperMimic (oral; UC Berkeley) learns one differentiable world model of the dynamics shared across a large motion library and trains the tracking policy through short imagined rollouts. After 3 hours on a single GPU it surpasses a baseline motion tracker trained for 7 days on 128 GPUs, a reduction of more than 1,000× in GPU-hours. It fails on 25 of 129,514 benchmark motions against 179 for the baseline, reaches zero failures after 9.8 hours on two GPUs, and transfers to a physical Unitree G1.
Play2Perfect (Stanford and Cornell) pretrains a dexterous hand through unstructured play in simulation, then fine-tunes on assembly with sparse rewards. The pretrained prior is 33× more sample-efficient than RL from scratch and transfers zero-shot to 60% success on insertions with 0.5 mm clearance. It uses no teleoperation and no human video.
What is left for data when simulation supplies the motion and compute stops being the constraint? Two things, both visible in these papers. GRAIL's authors chose known 3D configurations over reconstructing in-the-wild video. We read that choice as a statement about calibration: with camera and scale known, recovering metric 4D trajectories is a tractable step, and without them it is the step that fails. That applies to real capture as much as to virtual. And SuperMimic's world model holds its prediction error stable from 1 to 264 hours of motion data, which moves the question from how to train on a motion library to what the library covers.
Against the thesis, and we count it. GRAIL and Play2Perfect reduce the need for captured human demonstrations, while SuperMimic makes training from an existing human-motion library dramatically cheaper. If synthetic-first generation extends from loco-manipulation to dexterous household manipulation, the human-video conversion stacks above become a bridge technology, not a foundation. The papers do not show that extension yet, and all three still validate on real hardware.
Touch and driving: two quieter lessons
T-Rex (UC Berkeley and collaborators) puts touch inside the action loop. The authors collect a 100-hour tactile-rich dataset organized around basic motor primitives, process high-frequency touch signals at their own rate alongside the slower vision stream, and encode how contact changes over time instead of a static snapshot. On 12 manipulation tasks requiring delicate force control and deformable-object handling, average success is over 30% higher than the strongest baseline. The paper has company: at least a dozen accepted papers carry "tactile" in the title. For anyone maintaining a dataset schema, touch is no longer an optional column, and it arrives at a higher rate than the cameras it must stay synchronized with.
DriveMLP is the driving paper we would send to a planning team. Its Expert-Guided RL framework uses reinforcement learning purely as an optimizer, with a reward built from closeness to logged expert trajectories instead of handcrafted rules. A lightweight multilayer perceptron trained this way scores 86.51 on the nuPlan Val14 Reactive benchmark, 3.2 points above the previous best imitation-learning planner, with more than 15× faster inference. The authors' controlled study attributes a substantial share of the gain to the training procedure, not to model size. Our inference: when the reward is computed from logged expert trajectories, errors in the logs become errors in the reward.
The loop is closing on real robots
ENPIRE (NVIDIA, Carnegie Mellon University, and UC Berkeley) gives coding agents a harness for the physical research loop: reset the scene, execute a policy, verify the outcome, refine, and repeat. The authors report a 99% success rate on dexterous tasks such as organizing pins into a pin box and cutting a zip tie, and they test research teams of one, four, and eight agents. Two details deserve a careful read. The 99% is a pass@8 figure, meaning success within up to eight in-context retries per subtask, according to the project page. And the loop rests on the verify step, an automatic judgment of whether each rollout succeeded. That judgment is a label. An agent that improves a policy overnight is only as trustworthy as the verifier that scored its rollouts. The abstract does not report the verifier's error rate, and that is the number we would ask for first.
The surrounding program points the same way. Hang Zhao of Tsinghua University opens his early-career keynote abstract with a sentence worth keeping:
The Learn@Deploy workshop starts from the premise that "standard paradigms rely on static policies that cannot correct errors, self-critique, or optimize online." Roberto Martín-Martín's early-career keynote is titled "Learning Without Starting Over: Expanding What Robots Can Learn in the Real World." The attention is moving to what happens after deployment, which is the subject of the evaluation bottleneck: the robot that improves is the one whose failures are observed, scored, and routed back into a collection plan.
What we take from Austin
Stepping back from the individual papers, here is what we take from them.
- Budget by the labeled hour. EgoScale, Ego2Robot, and HuRo all converted human video before training on it. A capture plan without a conversion plan is half a budget.
- Give every demonstration an identity before asking what it is worth. Influence scoring, progress weighting, and anchor selection all operate on individual sequences. They need a stable handle on each one and a join to the evaluation that judged it.
- Record how the sequence went. Outcome, quality, speed, mistakes, and recoveries turned failed rollouts into training data in π0.7. Deleting failures now has a cost.
- Carry calibration and scale with the capture. GRAIL sidesteps the reconstruction problem by knowing its camera and scale in advance. Real capture can know them too, if they are recorded at the source.
- Record long, and put touch in the schema. RoboTTT rewards continuous sessions, and T-Rex is one of many papers treating touch as a first-class stream.
This is the ground Avala works on. Our capture programs use a versioned coverage catalog, whose Edition 4 contains 218 workflow specifications in 28 families. Every delivered sequence carries outcome labels, subtask boundaries, quality and speed grades, mistake and recovery segments, and rig and calibration IDs. Human and automated labeling runs across camera, LiDAR, and 4D reconstructions, and each dataset release has a durable identity, so that a curation method has something stable to score. The customer trains and owns the model. None of these papers evaluated our infrastructure, and we did not build it in response to them. The overlap is why we read them closely.
The strongest counter to our reading is in the same program. GRAIL and Play2Perfect avoid captured human demonstrations; SuperMimic makes better use of a human-motion library. WARP-RM and A Balanced Data Diet curate without human labels. ENPIRE automates verification. EgoScale's scaling law held on labels its own authors call noisy. A reasonable reader could conclude that the human share of this work shrinks every year. Our narrower claim is that the work moves instead of disappearing: from recording to converting, from labeling everything to labeling outcomes and exceptions, and from building datasets to deciding what goes into the next one. That claim is consistent with these papers. It is not proven by them, and it is the one that favors what we sell.
One calibration note before the close. Nearly every number above is an author-reported result, most from 2026 preprints that have not been independently reproduced, and several come from abstracts instead of full experimental tables. The official awards list had not been published when we wrote this, so we name GRAIL as a finalist on the strength of its authors' page.
Our bet coming out of Austin: as human video, simulation, and world models keep getting cheaper, the scarce asset will not be hours. It will be the record that says what each hour contains, how it was converted, how the sequence went, and which evaluation asked for it. That is the problem we have chosen. If you are working on it too, we would love to compare notes.
