Dexterity-BEV: Scaling up Dexterous Manipulation Learning with BEV Alignment
A unified 3D spatiotemporal representation aligns heterogeneous robot data and establishes a common data measure on the path to Physical AGI.
milestone
Milestones on the road to Physical AGI
We believe the key to Physical AGI is not merely larger models, but the efficient acquisition, generation, and amplification of high-value experience from the physical world.
Physical AGI requires intelligence to move from the digital world into the three-dimensional physical world, closing the loop from perception and future dynamics modeling to causal reasoning, action, and learning from real feedback.
DexForce is building the essential technology ladder for Physical AI through scalable real-world data, scalable generative data, and world-model-driven learning of future dynamics.
Through Dexterity-BEV, EmbodiChain, and DexWorldModel, we turn physical interaction experience into trainable, transferable, and deployable intelligent assets.
DexWorldModel: Causal Latent World Modeling towards Automated Learning of Embodied Tasks
Intelligence moves beyond perceiving the present to forecasting the future, learning physical dynamics and task evolution in a causal latent space.
EmbodiChain
Real data, generative simulation, model training, and deployment feedback form a closed loop that continuously scales and transfers physical experience.
DexVerse Simulation Engine
Real-time simulation of collisions, contact, deformation, and fluid dynamics, enabling high-fidelity virtual training and sim-to-real transfer.
Popcorn Scooping | Sim-to-Real
Adaptive scooping across simulated and real-world environments, demonstrating state understanding and Sim-to-Real capabilities.
Book Retrieval & Delivery | Sim-to-Real
Autonomous book grasping, retrieval, and delivery, with simulation-trained skills transferred to real-world operation.
Generative Simulation
Automatically generates diverse scenes, object configurations, and robotic tasks, enabling scalable, continuous production of embodied AI training data.
VR Teleoperation | Block Manipulation
VR-controlled block grasping, alignment, and placement, with real-time simulation of collisions, contact, and spatial alignment.
VR Teleoperation | Popcorn Scooping
VR-controlled block grasping, alignment, and placement, with real-time simulation of collisions, contact, and spatial alignment.
VR Teleoperation | Dominoes
Robot-driven domino chain reactions, demonstrating real-time simulation of rigid-body collisions, friction, and cascading physical interactions.
Open Dataset
Dexterity-BEV Dataset
3D-Space-UnifiedMultimodal Fusion Dataset
Foundational Data for Embodied Intelligence Training
Data Volume200TB
A large scale embodied interaction database spanning rich and diverse scene sampling, providing comprehensive data supply across all training stages of embodied models — from pretraining and fine-tuning to reinforcement learning.
3D Spatial Coordinate Alignment
A cross-embodiment dataset built upon the Dexterity-BEV paradigm, aligning multi-source, multi-view, multi-embodiment robotic data into an invariant, unified 3D canonical space and action representation.
Multi Source Heterogeneous Data
Covering multiple heterogeneous data production paradigms, this dataset systematically aggregates large scale source data from real robot teleoperation, simulation synthesis, and human demonstration.
Competition & Original Achievements
Cross-Dimension Intelligence Wins the ICRA WBCD Championship!
Cross-Dimension Intelligence claims the championship at the inaugural ICRA WBCD Bimanual Robot Challenge, emerging as the winner of one of the world’s leading robotics competitions.
