Core Technology

DexForce' s world-class team leads in AI, computer vision, generative 3D, and 3D perception large models. Frontier research drives our DexVerse™, AI, and synthetic data engine to realize general robot intelligence.

最新研究

DexWorldModel: Causal Latent World Modeling towards Automated Learning of Embodied Tasks

了解更多

EmbodiChain

一个自动化、模块化的具身智能平台,以更高效率扩展训练规模,加速具身 AI 学习的规模化进程

了解更多

DexGen Data

DexSim Simulation

DexAI Model

DexGen Data

- Research: 3D/4D data generation & characterization (3D reconstruction, unified geometric/physical representation, text-to-3D) supporting robot learning.

- Pioneered 3D/4D representation/generation precision-guaranteed, fast automated large-scale high-quality 3D data.

DexSim Simulation

- Research: High-fidelity virtual environments for robot learning’s massive interactive data needs (includes physics/rendering, efficiency, data augmentation).

- Quantity-to-quality leap: Scalable data augmentation + real-world-approaching perception & physics.

DexAI Model

- Research: Enabling learning and reasoning in robotic agents through end-to-end embodied intelligence, with a focus on Sim2Real transfer and multimodal large model integration.

- Via DexGen & DexSim' s efficient automated data pipeline, it establishes a full AI capability system: affordance understanding, natural interaction, and general manipulation.

Technological Achievements

Technological Achievements

Champion of the 1st WBCD Dual-Arm Robot Challenge

Top Prize at the World-Class Robot Competition

IROS Object Pose Estimation: Multiple championships/runner-ups

F-ConvNet/HotSpotNet: Top ranks on KITTI (autonomous driving benchmarks)

DualPoseNet: No.1 accuracy on NOCS (full-DOF pose estimation)

SRDC: Top precision on synthetic-to-real autonomous driving transfer benchmark

IEEE BTAS 2016: 1st in Video Person Recognition Evaluation

Intelligent visual rock-crushing system (Finnish quarries) – efficiency boost

GPNet et al.: deep learning unregistered object grasping + largest simulated dataset

World’s first 3D AffordanceNet: 3D affordance method + benchmark dataset (supports HRI R&D)

SkeletonNet/ToMoNet: first deep learning complex topological surface generation; CVPR 2019 Best Paper Candidate

Analytic Marching (lossless meshing theory/algorithm) – AnalyticMesh open-sourced

Adaptive deep model optimization: bridges MVS & deep reconstruction, enhancing MVS performance

Please fill out the form below. Our expert advisor will contact you shortly!
Inquiry:
Notes:
Cancel
Submit