Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
DynaRend: Learning 3D Dynamics via Masked Future Rendering for Robotic Manipulation
Authors: Jingyi Tian, Le Wang, Sanping Zhou, Sen Wang, lijiayi, Gang Hua
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate Dyna Rend on two challenging benchmarks, RLBench and Colosseum, as well as in real-world robotic experiments, demonstrating substantial improvements in policy success rate, generalization to environmental perturbations, and real-world applicability across diverse manipulation tasks. |
| Researcher Affiliation | Collaboration | 1National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, Institute of Artificial Intelligence and Robotics, Xi an Jiaotong University 2Amazon |
| Pseudocode | No | The paper describes methods in natural language and figures (e.g., Figure 2) but does not include explicit pseudocode or algorithm blocks. |
| Open Source Code | No | Code will be released in camera-ready version. |
| Open Datasets | Yes | We evaluate our method on two challenging robotic manipulation benchmarks, RLBench [21] and Colosseum [32]. |
| Dataset Splits | Yes | We collect 100 expert demonstrations per task for training on both benchmarks. For both settings, each task is evaluated over 25 rollout episodes. |
| Hardware Specification | Yes | Training is conducted using 8 NVIDIA RTX 3090 GPUs. All experiments are conducted using a Franka Research 3 robot arm equipped with a parallel gripper, mounted on a fixed tabletop setup. For visual input, we employ two calibrated Orbbec Femto Bolt RGB-D cameras |
| Software Dependencies | No | The paper mentions several software components such as 'Swi GLU', 'Ro PE', 'QK Norm', 'Franka ROS', 'Move It', 'kalibr package', 'easy handeye package', 'See3D', and 'Depth Anything v2', but does not provide specific version numbers for any of them. |
| Experiment Setup | Yes | During both pretraining and fine-tuning stages, we apply SE(3) augmentations to the input point clouds, camera poses, and action labels. Specifically, we perform random translations along the x, y, and z axes by up to 0.125 m, and random rotations around the z-axis by up to 45 degrees. The triplane grid is constructed with a resolution of 16 16 16. For pretraining, we train the model for 60k steps, while for fine-tuning on downstream tasks, we train for an additional 30k steps. In both stages, we use a batch size of 256 and set the initial learning rate to 1 10 4 with cosine decay schedule. Hyperparameters. We present the hyperparameters used in Dyna Rend as shown in Tab. 5. |