CRONOS: Benchmarking Non-Episodic Reinforcement Learning for Robotic Manipulation
CRONOS: Benchmarking Non-Episodic Reinforcement Learning for Robotic Manipulation
This paper introduces CRONOS, a simulation benchmark for evaluating non-episodic reinforcement learning (RL) fine-tuning of vision-language-action (VLA) models. The authors demonstrate that naive fine-tuning of pre-trained policies fails in non-episodic settings due to shifts in object configurations and end-effector poses. However, they show that these challenges can be mitigated through targeted reset allocation and addressing biases in pre-trained models. Furthermore, non-episodic training is found to enhance long-horizon manipulation and improve generalization to unseen object configurations and task sequences.
The authors introduce CRONOS, a simulation benchmark designed to evaluate non-episodic reinforcement learning for fine-tuning vision-language-action models, demonstrating that naive fine-tuning fails under object configuration shifts but targeted reset allocation and bias correction improve long-horizon manipulation and generalization.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- Environment
- SimulationOpen source S3
- code: not_reported
- weights: not_reported
- dataset: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
ned from scratch, its behavior when fine-tuning modern large-scale pre-trained vision-language-action (VLA) models remains underexplored. In this work, we introduce CRONOS, a simulation benchmark for RL fine-tuning of VLAs under non-episodic scene resets. CRONOS leverages high-fidelity physics simulators and adopts shared-scene multi-task settings. Using CRONOS, we show that naively fine-tuning pre-trained policies f
Open source S3
i-ntu.github.io/cronos/
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org