DexScale: Automating Data Scaling for Sim2Real Generalizable Robot Control
Authors: Guiliang Liu, Yueci Deng, Runyi Zhao, Huayi Zhou, Jian Chen, Jietao Chen, Ruiyan Xu, Yunxin Tai, Kui Jia
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate that the Dex Scale pipeline can be seamlessly integrated into the training of widely studied embodied robot policies... We demonstrate that Dex Scale can significantly advance the Sim2Real learning of deployable policies across multiple scenarios. 5. Experiments The pipeline of Dex Scale is generic and agnostic to simulation platforms... We evaluate the performance of Sim2Real deployment across various applications and assess the validity of Real2Sim projection in terms of action and object mapping by addressing the following questions: 1) Generalizability: How effectively does Dex Scale bridge the Sim2Real gap between simulated environments and real-world applications? 2) Scalability: Can the control skills learned by Dex Scale be scaled across different models and embodiments? 5.1. Generalizability: Bridging the Sim2Real Gap Experiment Setting. This experiment aims to quantify Dex Scale by how effectively it can overcome the Sim2Real gap... The success rates are presented in Table 2. Table 2: Success rates of imitation policies learned by different datasets under various Sim2Real gaps. |
| Researcher Affiliation | Collaboration | 1 School of Data Science, The Chinese University of Hong Kong, Shenzhen 2 Dex Force, Shenzhen. Correspondence to: Kui Jia <EMAIL>. |
| Pseudocode | No | The paper describes methods and algorithms but does not present them in clearly labeled pseudocode or algorithm blocks. The pipeline is illustrated graphically in Figure 2, and steps are described in prose. |
| Open Source Code | No | The project webpage at: https://edemai.github.io/dexscale.github.io/. This is a project webpage, not a direct link to a code repository, and the text does not explicitly state that the code is available there or elsewhere. |
| Open Datasets | Yes | Dex Scale supports retrieving objects in the scene from the Objaverse-XL dataset (Deitke et al., 2023) |
| Dataset Splits | Yes | A.3.1. OBJECT GRASPING Dataset Details. The dataset consists of 2000 trajectories for each feature and each setting (gap, domain randomization, domain adaptation, domain randomization and adaptation), with each trajectory containing 50 steps... The dataset is split into 99% training and 1% validation. A.3.2. OPEN BOX Dataset Details. The dataset consists of 50 demonstrations in the training set, with each demonstration containing 75 timesteps. The dataset is split into 50 demonstrations for training and 10 demonstrations for testing. Additionally, 100 evaluation trials are conducted in the simulator to assess performance. A.3.3. TABLEWARE REARRANGEMENT Dataset Details. The dataset used in this work consists of multiple trajectories designed to capture diverse scenarios for training and evaluation. Specifically: 1) Number of Trajectories: The dataset includes 200 trajectories, each consisting of 200 steps. |
| Hardware Specification | Yes | A.3.1. OBJECT GRASPING Training Environment. Training was conducted on 4 NVIDIA A800 GPUs, with each policy trained for 36 hours. A.3.2. OPEN BOX Training Environment. Training was performed on a single NVIDIA A100 GPU for a total of 48 hours. A.3.3. TABLEWARE REARRANGEMENT Training Environment. The training was conducted on a single NVIDIA A100 GPU using Py Torch version 2.0.1 as the deep learning framework. Table 4: Robot Specifications for Different Tasks Task Robot Name DOF Maximum Reach (mm) Maximum Payload (kg) Object Grasping Rokae SR3 6 705 3 Open Box AUBO I5 6 886.5 5 Tableware Rearrangement Widow X 250 S 7 650 0.25 |
| Software Dependencies | Yes | A.3.1. OBJECT GRASPING Training Environment. ...Py Torch version 2.0.1 was used as the deep learning framework. A.3.2. OPEN BOX Training Environment. ...Py Torch version 2.0.1 was used as the deep learning framework. A.3.3. TABLEWARE REARRANGEMENT Training Environment. ...using Py Torch version 2.0.1 as the deep learning framework. |
| Experiment Setup | Yes | A.3.1. OBJECT GRASPING Training Hyperparameters. The model is trained with a batch size of 12 for 200, 000 epochs using the Adam W optimizer with a learning rate of 1 10 5. The training objective is guided by the mean squared error (MSE) loss for actions and a 1-dimensional error for image latents. A.3.2. OPEN BOX Training Hyperparameters. The training process uses a batch size of 16 and runs for 150,000 epochs. The model is optimized using the Adam optimizer with a cosine learning rate scheduler, starting with a learning rate of 5 10 4. The training objective is guided by the Chamfer Distance loss function for point clouds and the Mean Squared Error (MSE) loss for action predictions. A.3.3. TABLEWARE REARRANGEMENT Training Hyperparameters. The training process utilizes a batch size of 8 and runs for 40,000 iterations. The model is optimized using the Adam optimizer with a cosine learning rate scheduler, starting with a learning rate of 1 10 4. The training objective is guided by the Mean Squared Error (MSE) loss function. |