Four bottlenecks, three workarounds: robot learning in 2026

The strongest robot learning papers of 2026 share a shape: each one attacks a long-standing bottleneck, and three of the four, read closely, are workarounds for datasets the field cannot collect yet. Long-horizon memory, embodied reasoning, dexterous manipulation, and efficient world modeling are the four problems that keep vision-language-action (VLA) models (policies that map camera observations and a language instruction directly to robot actions) stuck in demo mode. At Robotics: Science and Systems (RSS) 2026 and in the arXiv wave around it, four research lines mounted the most credible attacks on those problems we have seen.

Look at how each one gets there. One manufactures its memory supervision with large language models (LLMs). One bootstraps reasoning data that has no oracle source. One skips teleoperation entirely and learns dexterity from procedurally generated shapes. One treats video prediction as a training scaffold to be discarded at inference. We argued in the VLA data problem that beyond a capability threshold, data quality dominates architecture. This season reads like the field's strongest groups conceding the point: each workaround marks the location of data that is too scarce, too awkward, or too expensive to gather. The world-model line is the exception we count against ourselves, and we come back to it at the end.

Three numbers preview the story: task horizons, sample efficiency, and control latency are all moving at once.

Memory: MEM stretches task coherence to 15 minutes

MEM gives a VLA two memories, one visual and short, one textual and long, and stretches coherent execution to tasks around 15 minutes. A standard VLA is amnesiac: it sees the current frame and the instruction, and nothing else. That is fine for folding a shirt and fatal for cooking, where nothing in the current observation says how long the cheese has been melting or which cabinets were already searched. Memoryless policies loop, repeat finished subtasks, and cannot learn from their own failed attempts.

MEM (Multi-Scale Embodied Memory), from Physical Intelligence and academic collaborators, splits the problem by timescale. Short-term memory stays visual: a pretrained vision transformer gains interleaved spatial and causal temporal attention, history is aggregated into the current frame's tokens, and the past frames are then dropped before the VLA backbone runs. Attention cost falls from O(n²K²) to O(Kn² + nK²) for a K-frame history, which is what lets an 18-frame, multi-second window fit a real-time budget (under 380 ms per step on an NVIDIA H100 in the reported setup). Long-term memory becomes language: a high-level policy maintains a recurrent textual scratchpad ("ingredients gathered; cheese on pan; waiting for melt") and feeds it back as conditioning for the next decision, because a sentence of task state costs a small fraction of the tokens that raw frames would.

The demonstrations are long-horizon in a way the field has rarely shown on real hardware: recipe preparation, full kitchen cleanup, and timed cooking, on which memoryless baselines loop or forget. Memory also buys in-context adaptation. Success on opening a refrigerator with an unknown hinge side improves 62% because the policy can observe its own failed first attempt, and variable-height chopstick grasping improves 11%. On short-horizon dexterity MEM matches the π0.6 VLA it builds on, so the added memory is not taxing the reflexes, and it outperforms pooling-based and proprioception-only memory baselines.

Look at where the supervision comes from, though. The scratchpad is trained on LLM-generated summaries derived from subtask labels and success and failure signals. Someone had to record long sessions carrying exactly that annotation hierarchy: low-level motion, mid-level subtasks, high-level goals, and outcomes. That is the hierarchy we described in treating data collection as infrastructure, and MEM is what it is for.

Workaround, over a real architectural core. The two-timescale memory split is a genuine advance and would be worth having even with perfect data. What is substituting is the supervision: the scratchpad trains on LLM-generated summaries because no corpus of long-horizon robot sessions carries human-written task-state annotation. Record sessions labelled at motion, subtask, goal, and outcome level and the synthesis step becomes optional.

Reasoning: R&B-EnCoRe bootstraps data that has no oracle

R&B-EnCoRe shows that embodied reasoning data can be synthesized rather than collected, and that less reasoning, better chosen, beats more. Chain-of-thought helps VLAs, but current templates force exhaustive narration: every visible object, the plan, the gripper state, affordances, counterfactuals. Most of it is irrelevant to the next action, all of it costs latency, and there is no oracle dataset of good embodied reasoning to learn from instead. The chicken-and-egg is real: validating a reasoning trace requires a robust policy, and training a robust policy requires validated traces.

R&B-EnCoRe (RSS 2026, from Stanford and collaborators) breaks the loop by treating the reasoning trace as a latent variable sitting between context and action. A proposer samples diverse candidate traces by mixing reasoning primitives, with some randomly dropped. A validator then scores each trace with importance-weighted variational inference, and the weight that matters most is action predictiveness: how much the trace improves prediction of the expert's action. Concision and non-triviality enter as secondary pressures. Resampling produces a refined synthetic dataset of embodiment-aligned reasoning, the VLA retrains on it, and the cycle repeats, with no human reward engineering and no external verifier.

Across VLAs from 1B to 30B parameters the gains are consistent: roughly 28% relative success improvement on LIBERO-90 simulation and WidowX hardware with generated tokens cut about in half, a roughly 101% score improvement across navigation embodiments, and 21% lower collision-rate metrics in driving evaluations. The share of traces that mention task-critical objects rises from near zero to over 25%, and the learned reasoning is embodiment-specific: manipulation policies converge on contact points and gripper paths, locomotion policies on structural affordances, and driving policies prune hallucinated annotations. Because the method supports action forcing (dropping reasoning tokens at inference), the latency cost can go to zero while the training benefit stays.

The lesson generalizes past reasoning. When the data you need has no source, define what makes an example good (here, that it predicts expert actions), then let the model propose, score, and retrain. It is a recipe for manufacturing embodiment-aligned data, and we expect to see it reused.

Workaround that may beat the data it replaces. This is the one case of the four where we would not simply prefer the missing dataset. There is no oracle corpus of embodied reasoning, and it is not obvious that a human-written one would help. What the paper establishes is that the traces which improve action prediction are embodiment-specific; our inference, not its finding, is that this makes human-authored traces a poor fit, since nobody has published a human-versus-synthetic comparison to settle it. Synthesis here is not a second-best substitute for collection. It may be the right primary method.

Dexterity: SimToolReal replaces robot demonstrations with shapes

SimToolReal reaches zero-shot dexterous tool use without a single robot demonstration (no teleoperated trajectories and no action labels; a human video still specifies the task at deployment, as below), by reframing all of tool use as one goal-reaching problem. Dexterous manipulation is the data that teleoperation collects worst: puppeteering a 22-degree-of-freedom (DoF) hand through in-hand reorientation is unnatural for operators, force feedback is limited, and the mismatch between human and robot hands poisons direct retargeting. The economics compound the ergonomics: NVIDIA's Jim Fan puts effective teleoperation yield at roughly three hours per robot per day.

SimToolReal (RSS 2026, from Cornell and Stanford) trains one goal-conditioned policy entirely in simulation to move procedurally generated primitives (cylinders and cuboids of varying geometry and density) through sequences of random 6D goal poses, using massively parallel reinforcement learning across tens of thousands of simulated robots. The policy itself is small and fast: a long short-term memory (LSTM) network mapping proprioception, object pose, a grasp bounding box, and the goal pose to joint targets for a 22-DoF hand and 7-DoF arm at 60 Hz. Domain randomization includes external forces that knock objects out of the grasp, which trains recovery behavior as a side effect.

At deployment, a single human RGB-D (color plus depth) video is parsed by the Segment Anything Model (SAM) and the 6D pose tracker FoundationPose into a sequence of object goal poses, and the frozen policy tracks them. The video supplies goals only, never actions and never fine-tuning data. Zero-shot on 12 unseen real tools across 24 tasks and six categories (120 rollouts), the policy beats kinematic retargeting and fixed-grasp baselines by roughly 37% and matches specialist per-task policies. Drop the hammer mid-task and it reaches down, re-grasps, and resumes.

The follow-up, Play2Perfect, turns the same goal-reaching pretraining into a prior for precision assembly. Fine-tuned with sparse rewards on tasks derived from computer-aided design (CAD) models, it is 33× more sample-efficient than reinforcement learning from scratch and transfers zero-shot to about 60% success on insertions with 0.5 mm clearance, plus over 50% on multi-part assembly and screwing.

The honest accounting is in the failure modes: roughly 60% of SimToolReal's real-world failures are pose-tracking errors on small or symmetric objects under occlusion. The control problem yielded to simulation at scale. The bottleneck moved to perception ground truth.

Workaround that relocates the bottleneck rather than removing it. Procedural simulation genuinely replaces teleoperated demonstrations for the control problem, and that part is a real result. But roughly 60% of remaining real-world failures are pose-tracking errors, so the requirement moved from action labels to perception ground truth under occlusion. Demand for ground truth did not disappear, it changed type, from action labels to pose accuracy under occlusion. Whether the total falls we cannot say from these numbers: the 60% figure partitions the failures that remain, and we have seen no source comparing the before and after totals.

World models: imagination becomes a training objective

This summer's world action model papers converge on one claim: predicting the future is how you train a policy, not necessarily how you run one. We wrote in June about the shift from imitation toward world models; the new results sharpen that story.

DreamZero, from NVIDIA's GEAR Lab and collaborators, is the reference point. A world action model (WAM) jointly predicts future video and actions from a pretrained 14B video diffusion backbone, injecting physics and dynamics priors where VLAs inherit mostly semantic ones. The reported returns are large: 62.2% average task progress on seen tasks in unseen environments against 27.4% for pretrained VLA baselines, 49% on unseen verbs in DROID evaluations against 25 to 32% for the same baselines, cross-embodiment transfer from 10 to 20 minutes of video-only data yielding over 42% relative gains, adaptation to a new robot from about 30 minutes of play data, and roughly 7 Hz closed-loop control after heavy systems optimization.

Then came the correction, twice, and it turns on one design choice: whether the robot has to render the future before every move.

Fast-WAM ran the controlled experiment: removing video co-training hurts far more than removing future generation at inference. Keep the world-model objective during training, condition the action expert on the model's latents at deployment, and latency falls to about 190 ms per action chunk, 4× faster than the paper's own imagine-then-act ablation on matched hardware, at competitive success on the LIBERO and RoboTwin benchmarks. (That 4× is a within-paper comparison; it is not a ratio against DreamZero's 7 Hz, which comes from a different model and setup.) ImageWAM goes further, swapping video generation for an image-editing prior (predict what should change in the scene, not every frame) and conditioning actions on the editor's internal caches, at roughly a sixth of the compute and a quarter of the latency of its video-based counterpart. Latent-subgoal designs land in the same place: LaWAM conditions its policy on compact latent visual subgoals instead of rendered pixels and reports 98.6% on LIBERO at about 187 ms per action chunk. Edge-oriented variants report around 500 ms per 16-action chunk on Jetson-class hardware with 1B-scale models. The pattern is now common enough that a systematic survey of world action models organizes the field by exactly this choice, under the subtitle "Dream less, act more."

Dense pixel prediction, in other words, is becoming scaffolding: a way to force physical dynamics into the representation during training, then get out of the way at control time. That makes the data question concrete. If 10 to 20 minutes of the right video moves a policy across embodiments, then a calibrated, temporally coherent minute of physical-world capture has never carried more leverage.

Closest to a fundamental solution, on one axis only. Moving generation out of the control loop is a real efficiency result, not a data substitute, and it should survive any amount of new data. What it does not do is reduce the appetite for physical-world video: the training objective still requires it, and the cross-embodiment transfer results make a well-calibrated minute more valuable, not less.

Every workaround is a map of missing data

Step back and the four lines rhyme. MEM's memory supervision is manufactured by LLMs from subtask and outcome labels somebody recorded. R&B-EnCoRe synthesizes the reasoning dataset that could not be hired into existence. SimToolReal substitutes procedural simulation for demonstrations and one human video for task specification. The WAM efficiency papers exist because video is the one physical-world data type that is abundant while edge compute is not. In each case, the field looked at a data gap and engineered around it.

The honest counter to our own framing is that it risks being unfalsifiable. Any method that improves sample efficiency can be relabelled a workaround for data nobody collected, which would make "the field is data-limited" true by construction rather than by evidence. So here is the narrower version we will actually defend: in three of the four lines above, memory, reasoning, and dexterity, the authors name a specific dataset that does not exist and spend the method's central machinery manufacturing a substitute for it. The world-model line is the exception on that count, and we have marked it as one: moving generation out of the control loop is an efficiency result, not a data substitute. Note that reasoning qualifies on the count while still being the case where we would not simply prefer the missing data, which is why the two verdicts above read differently.

The papers also say, in their own failure accounting, where the engineering runs out. Pose tracking under occlusion is SimToolReal's dominant failure mode. Monocular video pretraining leaves depth imperfect: robots overshoot and undershoot targets even when the imagined future looks right, which is pushing world modeling toward 4D representations that carry depth and motion, not pixels alone. Tactile sensing, embodiment drift under long-term deployment (the slow change in a robot's dynamics as actuators wear and batteries age), and deformable and articulated objects sit on the open-problems lists across these works. None of those are architecture problems. They are ground truth problems: collection, calibration, synchronization, and labels with provenance, the same substrate training and evaluation both stand on.

The pieces also compose, and the composition has started shipping. Physical Intelligence's π0.7 release (April 2026) steers its VLA with visual subgoals generated by a lightweight world model, and conditions on episode metadata (speed, quality, and control modality) that makes even suboptimal autonomous data usable for training. NVIDIA has said the next generation of its GR00T humanoid models is built on DreamZero's architecture. A memory-augmented WAM with play-style motor priors and self-bootstrapped reasoning is a plausible 2027 system, and the groups involved have released code, weights, and benchmarks (DexToolBench among them, via the SimToolReal repository) that make the combination buildable. What that system will need is what each component already needs: long sessions carrying subtask and outcome labels, human video with reliable pose ground truth, metadata rich enough to grade demonstration quality, and enough embodiment and environment diversity that the priors survive contact with a real fleet.

One calibration note before the close: nearly every number above is an author-reported result from a 2026 preprint, consistent in direction across independent groups but not yet independently reproduced. With that stated, our bet from March stands: beyond a capability threshold, data quality dominates architecture. The 2026 crop does not weaken the claim. It shows the strongest groups conceding it, one workaround at a time, and it tells anyone building the next system exactly which data to take seriously. If you are working on any of these lines and thinking hard about the data underneath them, we would love to compare notes.