A robot-learning method that retrieves relevant past demonstrations while it is being tested was reported to achieve higher average success than comparison systems on held-out manipulation tasks. The paper reports an absolute gain of 17.60 percentage points on the simulated LIBERO benchmark and 20.83 points in a real-world UR5e robot setting. The authors also report that absolute success on out-of-distribution tasks, meaning tasks outside the training set, remained low.
A robot adapts by consulting examples
RA-VLA is built for in-context adaptation. It retrieves action-aware examples for a novel manipulation task and uses the retrieved context for grounded execution without updating its model weights at evaluation. The test protocol separated training tasks from evaluation tasks, with some tasks held out from training. That is the setting in which the study asks whether retrieval can support adaptation.
To select examples, the retrieval system looks for expert segments with similar behavior. The researchers align demonstrations from the same task using Dynamic Time Warping, or DTW, a trajectory-alignment method, then use those alignments to train the retrieval encoder with contrastive learning. Retrieval is based on cosine similarity.
A second training component, called contextual adherence loss, uses a regression margin to encourage action grounding in expert segments. The paper reports zero additional inference overhead for that loss. The architecture encodes each retrieved segment independently, and the authors report that this decouples inference latency from the scale of the retrieved context.
The test was held out by design
The study uses benchmark tasks, expert demonstrations and robot trials rather than a human-participant sample. LIBERO includes four task suites, each with 10 tasks and 50 expert demonstrations per task. The real-world UR5e evaluation includes four tasks and 30 expert demonstrations per task.
Training and evaluation were separated in both settings. For LIBERO, three of the four suites were used for training and the remaining suite was held out for evaluation; each held-out task was given three demonstrations and assessed in 50 trials. For UR5e, three of four tasks were used for training and one was held out, with four demonstrations and 12 trials per task. LIBERO data were also used as an auxiliary training source.
All evaluated frameworks used the same visual-language-action, or VLA, backbone, GR00t N1.5. The common setup used dual-view RGB observations, an action horizon of 16, four denoising steps and, by default, one retrieved expert segment.
The reported gap was visible in both settings
On the reported average success rates, the comparison baseline scored 20.9% on LIBERO and RA-VLA scored 38.5%. In the real-world UR5e evaluation, the corresponding figures were 35.4% and 56.3%. These were mean success rates over the stated task trials.
Another analysis used Relative Contextual Sensitivity, or Sctx, averaged across 1,000 samples. RA-VLA's score was 0.3639, compared with 0.0788 for RICLR and 0.0353 for a version without the adherence loss. The corresponding LIBERO-Goal success rates were 53.2%, 20.6% and 9.8%, respectively, leaving RA-VLA with the highest reported value in both measures among these comparisons.
In a separate retriever comparison on LIBERO-Goal, success was 10.2% with the baseline retriever and 53.2% with the action-aware retriever. The difference is an association within the reported trials, so the comparison alone does not establish that the retriever itself caused the higher result.
Retrieval speed was also measured. On an NVIDIA A100 GPU, retrieval took 0.18 milliseconds and accounted for 0.36% of total inference latency when the buffer held N = 107 stored keys. Under that reported benchmark condition, the paper says retrieval did not become a latency bottleneck as the buffer scaled.
On LIBERO-Goal, average success rates for buffer sizes N = 1, 2, 3 and 4 were 48.2%, 51.8%, 53.2% and 55.2%. For retrieval sizes K = 1, 2, 3 and 4, they were 53.2%, 54.0%, 54.8% and 54.6%. In the reported values, larger buffers were associated with higher success, while changing the number of retrieved segments produced less variation.
The result has a narrow boundary
The paper's authors report that performance depends on the quality and diversity of the demonstration buffer, absolute success on out-of-distribution tasks remains low, results are sensitive to hyperparameters and broader validation is needed.
The comparisons also come without reported inferential statistical tests or uncertainty calculations. The study does not report confidence intervals or standard errors for the success, contextual-sensitivity, retriever or buffer-size comparisons.
Taken together, the findings support a narrower reading: RA-VLA had higher reported performance in the stated held-out tests, but the evidence does not establish how well it would generalize beyond those settings. The authors call for broader validation.
The paper's status and support
The item is identified as an arXiv preprint dated 26 August 2026. Its supplied front matter also states Proceedings of the 43rd International Conference on Machine Learning, PMLR 306, 2026. The work was supported by Samsung Research Funding & Incubation Center of Samsung Electronics under project SRFC-IT2402-05.
Paper data and sources
Original title: RA-VLA: Retrieval-Augmented VLA for Test-Time Adaptation
Authors: Sanghwan Jang, Minjin Jeon, Minsoo Kim et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text