Zoellent AI Lab releases GR-SAC to refine frozen robot policies using a small controller
Read more
Pandaily
pandaily.com

Zoellent AI Lab releases GR-SAC to refine frozen robot policies using a small controller

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.

Similar stories

ACE Robotics introduces ACE-CaP-0: AI agents write reusable code for robots with 92.1% success rate
Read more
pandaily.com

ACE Robotics introduces ACE-CaP-0: AI agents write reusable code for robots with 92.1% success rate

ACE Robotics, a company focused on embodied artificial intelligence and led by co-founder SenseTime's Yang Xiaogang, has released the ACE-CaP-0 model. As part of this research approach, a general AI agent is responsible for writing, testing, and refining robot control software, which is then fixed as reusable skills that can operate autonomously without agent intervention.

The paper was posted on arXiv on November 12, and the company announced the work on November 13. This method is based on the 'Code as Policies' concept, which describes robot behavior through executable code instead of calling a large model at every step. The authors argue that multiple model calls slow down the robot, whereas the finished program utilizes the logic already embedded during its creation.

The main challenge lies in developing reliable code without requiring numerous expensive physical tests. In ACE-CaP-0, the agent independently decides when to request assistance.

Methodology and Results

In the comparison conducted by the team, the best results were achieved using on-demand demonstrations combined with selective simulation usage. The success metrics were obtained through the authors' own tests. Across eight real-world tasks, including charging station installation, bread packaging, bottle cap removal, cup insertion, flower placement, and shoe packaging, the recorded programs successfully completed 221 out of 240 attempts, achieving 92.1% success.

Seven of these tasks reached a success rate of at least 90%, and the charging station installation task demonstrated 100% success. The average task completion time ranged from 0.97 to 2.27 minutes. The programs function based on the robot's own perception and control, without human or coding agent intervention.

It should be noted that the results obtained are based on one setup with two manipulators selected by the team and have not yet been independently reproduced. ACE Robotics states that this work aligns with their plan to integrate general models with their global Kairos model.

D-Robotics and Infinigence AI integrate open robotics tools into Sunrise chips
Read more
pandaily.com

D-Robotics and Infinigence AI integrate open robotics tools into Sunrise chips

D-Robotics and Infinigence AI announced an agreement under which a set of open-source tools for embodied artificial intelligence will be integrated with D-Robotics' Sunrise chips and its RoboGo software platform. This was announced by both companies on November 13.

It is reported that some components are already working together, allowing developers to complete the full development cycle: training models, controlling devices, launching agents, and deploying them on robot hardware within a unified workflow.

This open-source toolset covers four stages of robot development. It was jointly developed by Tsinghua University and Infinigence AI.

D-Robotics is responsible for providing the hardware and platform. Its Sunrise chips provide access to devices and on-board operation execution, while RoboGo handles environment management, task planning, and execution monitoring.

According to company statements, the integration has already demonstrated results: developers can select a training environment in RoboGo, remotely control and call peripheral equipment, and conduct model training, agent execution, and testing with feedback from the hardware all within one process.

This announcement expands upon previous work shown by both parties. Infinigence AI opened the source code for APXInf in September, using the optimized Pi0.5 model for NVIDIA Jetson Thor, and the S600 adaptation added a domestic robot chip to its target solutions. D-Robotics previously combined third-party robot models with Sunrise S600, including GigaBrain-0.7 from GigaAI.

Neither company has published performance metrics for the combined workflow nor stated when more Sunrise models will support APXInf.

Ant Group's Robbyant releases LingBot-Touch model for tactile perception of robotic hands
Read more
pandaily.com

Ant Group's Robbyant releases LingBot-Touch model for tactile perception of robotic hands

Robbyant, a subsidiary of Ant Group, has released the LingBot-Touch model to the public. This fundamental tactile model is designed to provide robotic hands with a unified sense of touch.

Announced on November 10, the model was pre-trained on data collected from nearly 5000 hours of real tactile experience. It covers three sensory principles: magnetic, visual-tactile, and piezoresistive. The model code is available under the Apache-2.0 license, and the Tiny and Small checkpoints are hosted on Hugging Face and ModelScope.

The main problem that LingBot-Touch solves is the incompatibility of raw signals generated by different tactile sensors; touch experience gained from one type of sensor is rarely applicable to another. LingBot-Touch addresses this by providing each sensor type with its own input tokenizer, which feeds data into a common Transformer. The model maps contact points into a single fingertip coordinate space and outputs unified tactile tokens that can be used by various action models.

The repository describes a target approach in the JEPA style, which predicts latent representations from the spatio-temporal context of touch. Training data was obtained from both human-worn tactile gloves and teleoperated dexterous hands.

Testing the model on robotics tasks

Robbyant tested the model on two dexterous hands—Sharpa Wave and XHand—across 26 tasks performed by real robots. These tasks included turning pages, twisting bottle caps, and identifying Mahjong tiles. When adding the LingBot-Touch representation to four action models—LingBot-VLA, $\pi$0.5, Diffusion Policy, and T-Rex—the average task success rate increased by approximately 14–26 percentage points compared to versions without tactile feedback.

In one instance, when T-Rex attempted to twist caps on Sharpa Wave, the success rate was 56.25% without tactility, 62.5% using raw sensor signals, and 90% after integrating LingBot-Touch.

All 26 tests were conducted using the Tiny version, which has about 5 million parameters and increases the size of connected policies by only 0.13%–0.16%. The team notes that the larger Small version continues to improve as the volume of training data increases to 5000 hours.

To verify knowledge transfer across different hardware platforms, the team froze the encoder on XHand, whose magnetic sensing module was absent in the pre-training data, and trained only the action model on XHand demonstrations. Knowledge of magnetic touch in this case was acquired exclusively from human glove data.

Robbyant positions this model as a neutral layer between tactile sensors and action models, rather than a component tied to its proprietary LingBot-VLA. This allows hardware manufacturers to choose sensors based on cost, while system teams can retain their own policies.

All presented results belong to Robbyant itself, and the company has not reported independent reproduction of this data. The release on GitHub is an inference release containing the frozen encoder and utilities for extracting tactile features.

Popular