Zoellent AI Lab has introduced GR-SAC, a lightweight reinforcement learning-based controller designed to correct the operation of a frozen model that integrates vision, language, and actions. The method was announced in November in China, following the publication of the repository and the submission of results to the RoboMME leaderboard earlier that same week.
GR-SAC stands for Gated Residual Soft Actor-Critic. Instead of retraining the entire base model, the implementation adds a visual actor with 232,110 parameters to SimpleMemVLA. The original model continues to interpret observations and generate action sequences, while the smaller controller learns limited adjustments to the manipulator joint commands.
The controller analyzes images from frontal and wrist cameras, the robot's state, and the queue of pending actions. The inclusion of a gate allows corrections only when specific observable conditions are met, including relatively stable target actions and grasping commands. Constraints, decay, and resets control the accumulation of these corrections. These parameters are design solutions for control, not a guarantee against unsafe behavior in all environments.
In the presented benchmark results, the success rate increased from 706 to 711 episodes out of 800, corresponding to an increase from 88.250% to 88.875%. Consequently, the number of failures decreased from 94 to 89, representing a relative reduction in failures of 5.32%, rather than an increase in overall success by 5.32 percentage points.
Testing covered 16 RoboMME tasks, each consisting of 50 episodes. The InsertPeg task improved from 46% to 52%, and MoveCube increased from 94% to 100%. However, not all tasks showed improvement: the benchmark table also recorded regressions, so the aggregate result cannot be considered a universal gain across all manipulation skills.
Zoellent describes GR-SAC as a leader in the standard RoboMME track, which evaluates methods based on open models. The same combined leaderboard includes a separate Agentic category, which uses proprietary state-of-the-art models and human benchmarks. Therefore, a position in the standard track is not equivalent to superiority over every entry in the overall leaderboard.
The public repository contains a technical report, a checkpoint of the benchmark controller, training and evaluation materials, as well as results for each task. The code is licensed under MIT; this does not automatically waive the terms of the base model or other dependencies. The reported results remain the developers' benchmark evaluations, not an independently reproduced reliability test in a real production setting.



