REX: GPU-Accelerated Sim2Real Framework with Delay and Dynamics Estimation
Authors: Bas van der Heijden, Jens Kober, Robert Babuska, Laura Ferranti
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate our approach on two real-world systems, demonstrating its effectiveness in improving sim2real performance by accurately modeling both system dynamics and delays. Our results show that the proposed framework supports both accelerated simulation and real-time processing, making it valuable for robot learning. |
| Researcher Affiliation | Academia | Bas van der Heijden EMAIL Cognitive Robotics Delft University of Technology Jens Kober EMAIL Cognitive Robotics Delft University of Technology Robert Babuska EMAIL Cognitive Robotics, Delft University of Technology, CIIRC, Czech Technical University in Prague Laura Ferranti EMAIL Cognitive Robotics Delft University of Technology |
| Pseudocode | Yes | Algorithm 1: UKF Estimation Step with Delay Compensation Algorithm 2: CMA-ES (Covariance Matrix Adaptation Evolution Strategy) |
| Open Source Code | Yes | The documentation, tutorials, and our open-source code can be found at https://bheijden.github.io/rex/. |
| Open Datasets | No | The paper describes collecting data from real-world systems (pendulum and quadrotor) using specific sensors (Real Sense d435i camera, motion capture system). It does not explicitly state that this collected data is made public, nor does it provide any links or citations to publicly available datasets that were used. |
| Dataset Splits | No | The paper mentions comparing reconstruction error with "validation data obtained from the pendulum’s encoder" and performs "Zero-shot evaluations on the real system" after training policies in simulation. However, it does not provide specific details on how the collected real-world data itself is split into training, validation, or test sets in a reproducible manner, nor does it refer to standard splits for any public datasets. |
| Hardware Specification | Yes | We evaluated simulation speeds using the COMPILED runtime on an NVIDIA RTX 3070 Laptop GPU. |
| Software Dependencies | No | The paper mentions leveraging "JAX (Frostig et al., 2018)" and algorithms like PPO and CMA-ES, as well as referencing implementations such as "Clean RL" (Huang et al., 2022). However, it does not provide specific version numbers for JAX or any other key software dependencies, which are necessary for full reproducibility. |
| Experiment Setup | Yes | Appendix E, titled "Covariance Matrix Adaptation Evolution Strategy," provides a table summarizing the hyperparameters used for the CMA-ES algorithm for both Pendulum and Quadrotor tasks, including Generations, Population size, Learning rates, and other specific parameters. Appendix F, titled "Proximal Policy Optimization," similarly provides a table summarizing hyperparameter settings for PPO, including Total timesteps, Learning rate, Number of environments, Number of epochs, Discount factor, and other detailed settings for both tasks. |