A new preprint reports that a vision-language model was used to guide a robot-learning system that achieved higher reported success while collecting shorter episodes, according to experiments on a 22-task manipulation benchmark. The work presents EXIMO as a way to adapt a pretrained vision-language-action policy to new tasks without additional teleoperation, or human-operated demonstrations.
The central comparison was between the original policy and a version in which a vision-language model orchestrated exploration. The orchestrated version had a higher success rate with similar time-to-success, and its shorter episodes were reported as evidence of greater efficiency in gathering training data. The paper does not give exact numerical effect sizes for those comparisons in the text; the results are presented graphically.
A three-stage route from exploration to refinement
EXIMO is organised around three stages: explore, imitate and optimise. The starting point was the 3B version of Gemini Robotics On-Device, known as GROD. In plain terms, the system first uses a vision-language model to help search for useful behaviour, then trains the robot policy on the resulting demonstrations, and finally applies reinforcement learning to refine it.
The exploration stage was not simply a matter of letting the policy run unchecked. Successful rollouts were kept in a data buffer only when the environment’s ground-truth detector marked them as successful. During imitation training, the policy was conditioned on the actual goal while predicting actions. The resulting dataset therefore depended on which runs the detector accepted.
For the main exploration comparison, both methods were evaluated for 1000 episodes. The study then examined what happened when the base policy was fine-tuned on the orchestrated data and when that fine-tuned model received further online reinforcement learning. The optimise stage used residual off-policy reinforcement learning to further fine-tune the policy.
Fine-tuning carried the gains into the standalone policy
The reported advantage was not limited to the model acting with the vision-language model in the loop. After supervised fine-tuning on orchestrated data, the base vision-language-action policy was reported to perform markedly better than before and to outperform the orchestrated agent in the comparison. This was the paper’s key distillation result: behaviour discovered with external guidance was transferred into the base policy.
After online reinforcement learning, both models had better performance. The GROD model that had first undergone supervised fine-tuning began the online-RL comparison at a higher level and also converged above the unfine-tuned GROD baseline. The online-RL figure reports averages across seeds as means with standard-error uncertainty, so the curves represent an average rather than a single run.
Across the evaluated tasks, the model that received both supervised fine-tuning and reinforcement learning outperformed the other two baselines reported in that comparison. Taken together, the results suggest a staged pattern: language-guided exploration supplies useful experience, imitation transfers that experience to the base policy, and reinforcement learning can improve the resulting model further.
The results are encouraging, but the picture is not uniformly better
A supplementary experiment complicates the overall result. Residual-policy distillation improved during offline learning, but after the switch to online training it learned significantly more slowly than pure online reinforcement learning. That finding means the route into online learning matters: an offline improvement did not automatically translate into faster subsequent adaptation.
The paper’s reported comparisons are mostly graphical, and the text does not provide exact numerical effect sizes for several headline results. The online-RL plot includes seed-averaged means with standard-error uncertainty, while the supplementary residual-policy analysis reports averages with standard deviation. No formal hypothesis tests or p-values are reported in the supplied analysis.
What the preprint has not yet established
The evidence is confined to the reported manipulation-task benchmark and the tested GROD-based configurations. The experiments were conducted in simulation, and the method relied on a ground-truth success detector to decide which rollouts were successful. Those conditions leave open whether the reported gains would carry over to physical robots, different task distributions or a setting in which success must itself be judged by a vision-language model.
The study also does not establish that the system can reset an environment by undoing a task, or that the approach will work with on-policy distillation methods. Those are proposed next questions rather than results demonstrated by the current evaluation. The supplied metadata lists the work as an arXiv preprint dated 20 August 2026, with identifier arXiv:2608.19891.
Paper data and sources
Original title: EXIMO: VLM Guided Exploration of VLA Policies
Authors: Bhavya Sukhija, Oliver Groth, Mohit Shridhar et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text