Robot policies are learning to reuse demonstrations instead of demanding more

Success is not enough

Robot policy evaluation often relies on a binary success signal for task completion, according to a preprint on quality-controllable policies. A robot either completes the task or it does not, and that single bit says nothing about how it moved: too fast, too rough, too slow for the line it works on. For real deployments, that missing information is the difference between a demonstration that works in the lab and a policy that works on a factory floor. m:…4621

The preprint introduces Quality-Controllable Policy, or QCP, which defines desired execution behaviors through quality attributes such as execution speed and motion smoothness. QCP enables real-time adjustment of robot execution quality and allows inference-time control over individual quality attributes and their combinations, so an operator can ask for faster or smoother execution without retraining a new policy for each setting. m:…4621

The mechanism is an iterative hindsight relabeling loop that expands the controllable quality range. QCP collects online rollouts under target quality labels and fine-tunes the policy, which means existing robot experience gets relabeled with the quality it actually demonstrated, rather than discarded because it was not the exact behavior originally requested. Experiments on Robomimic and RoboEval demonstrate that QCP improves task execution quality and success rates, the authors report. m:…4621

The examples in the preprint are concrete. Controlling speed increases throughput in conveyor operations, and controlling transport stability improves success when carrying wine glasses. Both are cases where the same task needs different behavior in different situations, and where collecting a fresh demonstration set for each variant would be slow and expensive. m:…4621

Faster reuse of what the model already knows

A second preprint, OSWAM, applies the same reuse logic to action generation itself. OSWAM is a world action model for autonomous driving and embodied control that enables one-step action generation, according to the authors. Direct-policy world action models typically require tens of denoising steps, and reducing those steps in standard models degrades action quality, so speed has usually cost accuracy. m:…2775

OSWAM uses structured imagination tokens to initialize the diffusion process and privileged information distillation with future conditioning, the preprint states. The result is approximately 3x speedup over direct-policy baselines, with 90.3 PDMS on NAVSIM and 90.00% average success on RoboTwin 2.0. RoboTwin inference reaches 40.81 milliseconds on an NVIDIA RTX 5090, and the model was validated on real-world PiPER manipulation tasks. m:…2775

The pattern here is distillation rather than new data: knowledge the training process already had, including privileged information about the future, is compressed into a model that can act in a single step. That is reuse at the level of the model's own training signal, and it matters because robots cannot wait through dozens of denoising steps when they need to react. m:…2775

Routing the policies you already have

A third preprint, REA, reuses policies at run time instead of replacing them. REA unifies deliberative and reactive policies in a single runtime with asynchronous coordination, operating under the rule that no running policy waits on the brain, meaning slow high-level reasoning never stalls real-time control. It employs a route memory that stores each policy's measured success ahead of time, a boundary-triggered judge gate that consults the brain on policy execution, and a shared-budget escalation ladder for bounded recovery. m:…1545

The reported results are strong: 95.9% success on 3,000 LIBERO-Pro test episodes, against 94.7% for the best single-agent baseline in the policy pool, with control steps reduced by 25% and control time by 17% compared with baselines. In a 100-episode RoboTwin 2.0 study, REA reached 84% success with vision-language-action policies and 86% with vision-language-action plus code, against 80% for the strongest single baseline, MEDex. REA also transfers to the BEHAVIOR-1K environment and to a real Unitree G1 humanoid, the authors report. m:…1545

Taken together, the three preprints describe one shift. QCP relabels existing rollouts to cover more behaviors, OSWAM distills existing knowledge into faster action, and REA routes among existing policies instead of betting on one. In each case the gain comes from using what is already available more cleverly, not from a larger demonstration set. m:…4621 m:…2775 m:…1545

What is still missing

A benchmark preprint released alongside this work shows how far perception still lags. SOMA, a benchmark for fine-grained egocentric body-motion understanding, comprises 1,632 human-verified questions across body configuration, motion dynamics, and temporal reasoning. Humans achieve 68.5% accuracy from video alone, while the strongest multimodal model tested, Qwen3.6-27B, reaches 50.0% zero-shot accuracy, and models often misread the wearer's own body or misjudge events in time. m:…0385

That gap matters for the reuse story. Relabeling, distillation, and routing all assume the system can tell what actually happened in a rollout or a video. If body-motion understanding is still well below human accuracy, the labels being reused may themselves be noisy, and the gains reported on benchmarks may shrink on real hardware. The researchers evaluated 37 leading models and are releasing a training dataset and a multimodal baseline, which gives the field a way to measure whether that gap closes. m:…0385

For developers, the practical lesson is to instrument quality, not just success. Record speed, smoothness, and stability alongside completion, keep the rollouts that failed at one quality setting but succeeded at another, and route among policies by measured performance. Those are the habits that make relabeling and routing possible, and they cost far less than collecting a new demonstration set for every behavior variant. m:…4621 m:…1545