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..
Taccel: Scaling Up Vision-based Tactile Robotics via High-performance GPU Simulation
Authors: Yuyang Li, Wenxin Du, Chang Yu, Puhao Li, Zihang Zhao, Tengyu Liu, Chenfanfu Jiang, Yixin Zhu, Siyuan Huang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through extensive validation in object recognition, robotic grasping, and articulated object manipulation, we demonstrate precise simulation and successful sim-to-real transfer. These capabilities position Taccel as a powerful tool for scaling up tactile robotics research and development... Comprehensive evaluations (Sec. 5) demonstrate its characteristics: Precision: Taccel leverages advanced solid material simulation techniques (IPC [32] and ABD [31]) for physical accuracy. Scalability: ... achieves unprecedented parallelization. On a single H100 GPU, it reaches over 900 FPS in total (4096 environments, 18ˆ wallclock time) for a peg-insertion task... |
| Researcher Affiliation | Academia | 1 Peking University 2 University of California, Los Angeles 3 Beijing Institute for General AI 4 State Key Lab of General AI 5 Beijing Key Laboratory of Behavior and Mental Health, Peking University 6 Embodied Intelligence Lab, PKU-Wuhan Institute for Artificial Intelligence |
| Pseudocode | No | The paper describes the mathematical foundations of the IPC simulation framework in Section 3 and the unified ABD-IPC algorithm in Appendix A. While it references algorithms like DFC, it does not contain a dedicated pseudocode or algorithm block within the document. |
| Open Source Code | No | To foster community development, we will release Taccel codebase and documentations, while maintaining active collaboration with researchers to incorporate feedback, add features, and expand capabilities, ensuring its evolution as a comprehensive tool for tactile robotics research. |
| Open Datasets | Yes | We press a calibrated Gel Sight-type sensor perpendicullarly on 18 objects from a standard tactile shape testing dataset [16] on a real-world setup (Fig. 3(a))... We generate grasps on 10 diverse objects from Contact DB [4], YCB [7], and adversarial object [42] datasets. |
| Dataset Splits | Yes | We collect a training dataset 4K tactile depth maps of a tactile sensor pressing on them within Taccel, each with object pose and tactile depth randomly sampled... We split the samples into a training set (85%) and a validation set (15%). The former is used to train the classification model... |
| Hardware Specification | Yes | On a single H100 GPU, it reaches over 900 FPS in total... With a single NVIDIA H100 80G GPU, we benchmarked Taccel against SAPIEN-IPC. Noteworthy, while Taccel s FP64 precision is ideal for HPC GPUs (which have a high 1:2 FP32:FP64 FLOPS ratio), it remains highly performant on more accessible consumer cards like RTX 3090 and 4090; see Sec. C.2 for more details. |
| Software Dependencies | No | Taccel provides intuitive Python APIs designed to make tactile robotics simulation accessible to researchers while maintaining high performance through NVIDIA Warp [41]. While it mentions Python and NVIDIA Warp, it does not provide specific version numbers for these or any other key software components used in the methodology. |
| Experiment Setup | Yes | We trained a Res Net-18 model [20] for 10-category object classification using the simulated depth images... supervised by the NLL Loss, using the Adam optimizer with a learning rate 1e-4 for 100 epochs. During training, we apply an exponential LR scheduler. |