A robot controller can use recent task-progress estimates as a short memory before choosing its next movement. In a new method called Value as Memory (VaM), the researchers report that mean success across three physical-robot tasks was 29.4 percentage points higher than for a version without that progress history, averaging results from the final three saved model checkpoints.
The method addresses a problem with camera-based control: similar-looking views can occur at different stages, or after a robot has stalled or lost contact. VaM’s value model estimates how far a task has advanced from recent camera observations. The controller receives several of those recent estimates along with its visual and joint-position input, so a rising, flat or falling sequence can provide context that one image cannot.
The researchers tested box assembly, laundry folding and magnetic-block assembly on a physical robot. In laundry folding, for example, the reported success rate was 88.9% with value caching and 48.9% without it across the final three online-learning checkpoints. Success meant completing each task within its set time limit. The 29.4-point mean averages the three task results; it is not a measure of performance across many robot types or household settings.
After expert demonstrations and an offline-learning warmup, VaM has two online stages. Online DAgger collects attempts with human corrections; the later Learn from Experience (LfE) stage trains on autonomous attempts without new corrections, while retaining earlier intervention data.
In a separate comparison, the online schedule, which includes human-correction data collection and LfE, improved mean success by 32.8 percentage points over offline reinforcement learning after LfE, with matched gradient-step and data-collection budgets. It also reduced recorded experimental wall-clock time per gradient step by 57.8% on average versus offline training: online data collection and training ran concurrently, while offline RL collected data first and then trained. These are results for the study schedules and tasks, not a general estimate of deployment costs.
The paper is an arXiv preprint submitted on October 15, 2026. Its authors say the framework currently uses a single robot and depends on progress estimates that closely track the task; they identify offline learning as necessary for stable online updates. The study does not establish how the method would scale across a robot fleet or work beyond these three laboratory tasks, and its code was still being organized for release.
Sources
- Value as Memory: Online Value-Guided Post-training for General Robot Policies, arXiv:2610.22368v1, submitted October 15, 2026.
- Authors’ VaM project page and experiment overview.
- Illustrative photograph by Nenad Stojković (Shixart1985), via Wikimedia Commons, licensed under CC BY 2.0; not an image from the study.
