A method designed to preserve visual information in robot-control AI reported higher manipulation success than the GR00T N1.6 base model across three simulation benchmarks and two tasks on an AGIBOT A3 Ultra robot. Called V-Link, the method was tested on LIBERO, LIBERO-Plus and RoboTwin 2.0, as well as autonomous power-on and power-off tasks.
The clearest real-world gap appeared on the power tasks. V-Link succeeded in 98% of power-on trials and 94% of power-off trials, compared with 78% and 70% for GR00T N1.6. The comparison covered 50 consecutive trials per task.
The visual handoff is the target
V-Link addresses a handoff inside a vision-language-action system. A vision-language model, or VLM, processes visual and language information, while an Action DiT receives transferred features to generate robot actions. The design separates spatial information about geometry from semantic information about the scene, then makes both available to the action model.
To test whether that information remained accessible, the researchers froze the VLM and Action DiT features and trained lightweight depth-estimation and semantic-segmentation heads under the same supervision. In the GR00T N1.6 base model, Action DiT features produced a depth MAE of 0.071, compared with 0.015 from VLM features. Their segmentation mIoU was 0.290 versus 0.665. The diagnostic result means the action features were less useful than the VLM features for recovering both scene geometry and semantic labels.
The largest differences appeared under shifted conditions
On LIBERO, V-Link ranked first across all four suites, with a 99.3% average success rate. The reported gain over GR00T N1.6 was 1.9 percentage points, rising to 5.1 points on LIBERO-Long.
On LIBERO-Plus, V-Link recorded 75.0% overall success across seven distribution-shift settings and ranked first in six. The reported gain over GR00T N1.6 was 31.2 points. The reported differences were especially large under noise, at 60.4 points, language, at 47.3 points, and lighting, at 30.6 points.
RoboTwin 2.0 gave V-Link a 56.8% average across six selected tasks, 18.8 points above GR00T N1.6, with higher reported results on every task. The joint training setup used 50 demonstrations per task, or 300 in total, and the benchmark evaluation used 100 trials per task.
Two queries, one action stream
That performance came with a specific design. V-Link adds Spatial and Semantic Query tokens to the VLM, uses query-isolating attention and auxiliary depth and semantic-segmentation heads during training, then injects the learned queries into Action DiT through asymmetric pathways. Ground-truth depth and segmentation information is used during training but is not required at inference.
The diagnostic ablation separated the roles of the two queries: Spatial Queries had the best depth-estimation result, while Semantic Queries had the best segmentation result. V-Link Action DiT outperformed GR00T Action DiT on both tasks.
In the auxiliary-query comparison, average success was 38.0% for the baseline, 51.2% with Spatial Queries and 47.8% with Semantic Queries. Using both reached 56.8%, the highest reported average.
The combined supervision and injection condition also had the highest reported result. The baseline average was 38.0%; query injection without task supervision produced 40.3%, while task supervision without injection produced 39.8%. With both, the average was 56.8%, or 16.5 and 17.0 points above the two single-component conditions.
V-Link's reported inference latency increased by 1.58 milliseconds, from 43.21 to 44.79 milliseconds.
A focused test
The real-robot comparison used 100 teleoperated demonstrations per task for training and 50 consecutive trials per task for evaluation. In simulation, success rates were based on 50 trials per task on LIBERO and LIBERO-Plus and 100 on RoboTwin 2.0.
The supplied document is an arXiv preprint, version 1, dated 26 August 2026. Its evidence covers three simulation benchmarks, selected benchmark tasks, frozen-feature diagnostics, ablations and two AGIBOT A3 Ultra tasks. The results are therefore a focused report of performance in the tested settings, not a general verdict on every robot or task.
Paper data and sources
Original title: V-Link: Recovering Lost Visual Representations in Action DiT for Vision-Language-Action Models
Authors: Yehao Lu, Jiarui Yang, Yuning Su et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text