Home
Li Feifei’s Team Unveils Method to Generate Infinite Robot Training Grounds From Single Video
In embodied intelligence, bridging the gap from simulation to reality—known as "Sim2Real"—has long been a persistent hurdle. Simulation setups are expensive, and models frequently lose performance when deployed in physical settings. A recent collaborative study from Li Fei-Fei’s team, NVIDIA GEAR Lab, and Georgia Tech introduces a novel solution: Real2Sim.
Named "SimFoundry," this system creates interactive, trainable, and testable robotic simulations using simple real-world video. Going beyond traditional 3D reconstruction, SimFoundry performs deep scene analysis to automate environment creation, drastically reducing the effort required to build simulations.

SimFoundry’s core innovation involves a closed loop between "digital twins" and "digital cousins." The system extracts geometric, physical, and interactive data from videos to build high-fidelity digital twins. It then automatically varies appearance, layout, and tasks to generate diverse digital cousins. This allows developers to derive virtually unlimited training data from a single video, enabling end-to-end strategy learning and evaluation within the simulation.
Experiments confirm SimFoundry’s strong predictive power, with simulated robot performance closely matching real-world outcomes. Crucially, strategies trained on this synthetic data transfer effectively to reality without fine-tuning ("zero-shot"), excelling in complex tasks like multi-step operations and dual-arm coordination.
Co-authored by leading robotics experts, including key members of NVIDIA GEAR Lab and Li Fei-Fei’s team, this open-source system promises to transform embodied AI development. By replacing costly manual data collection with generative technology, robots can now transition from lab to real-world applications much faster.
Paper link: https://arxiv.org/pdf/2606.28276v1
Related article
Anthropic Launches Claude Mythos to Solve Erdős Conjecture With Concise Proof
On May 26, Sholto Douglas, an engineer at Anthropic, took to X to announce that their latest model, Claude Mythos, had cracked a century-old puzzle in combinatorial geometry: the Erdős unit distance conjecture. Dubbed a "clever and concise proof," th
Amazon to Acquire Decart for $7B, Securing Key AI Chip Adapter
Per the Wall Street Journal, Amazon is negotiating to purchase AI startup Decart at an estimated $7 billion valuation. If no competing bids arise, the transaction could close shortly, underscoring major tech firms’ urgency to secure computing resourc
How Locus tackles one of robotics' biggest challenges: manipulation
NeuraGrasp handles diverse items, including small containers, porous polybags, and textiles. | Source: Locus RoboticsLocus Robotics Corp. has solved numerous robotic hurdles in live operational settings. Its autonomous mobile robots (AMRs) have trave
Related Special Topic Recommendations
Comments (0)
0/500
In embodied intelligence, bridging the gap from simulation to reality—known as "Sim2Real"—has long been a persistent hurdle. Simulation setups are expensive, and models frequently lose performance when deployed in physical settings. A recent collaborative study from Li Fei-Fei’s team, NVIDIA GEAR Lab, and Georgia Tech introduces a novel solution: Real2Sim.
Named "SimFoundry," this system creates interactive, trainable, and testable robotic simulations using simple real-world video. Going beyond traditional 3D reconstruction, SimFoundry performs deep scene analysis to automate environment creation, drastically reducing the effort required to build simulations.

SimFoundry’s core innovation involves a closed loop between "digital twins" and "digital cousins." The system extracts geometric, physical, and interactive data from videos to build high-fidelity digital twins. It then automatically varies appearance, layout, and tasks to generate diverse digital cousins. This allows developers to derive virtually unlimited training data from a single video, enabling end-to-end strategy learning and evaluation within the simulation.
Experiments confirm SimFoundry’s strong predictive power, with simulated robot performance closely matching real-world outcomes. Crucially, strategies trained on this synthetic data transfer effectively to reality without fine-tuning ("zero-shot"), excelling in complex tasks like multi-step operations and dual-arm coordination.
Co-authored by leading robotics experts, including key members of NVIDIA GEAR Lab and Li Fei-Fei’s team, this open-source system promises to transform embodied AI development. By replacing costly manual data collection with generative technology, robots can now transition from lab to real-world applications much faster.
Paper link: https://arxiv.org/pdf/2606.28276v1
Anthropic Launches Claude Mythos to Solve Erdős Conjecture With Concise Proof
On May 26, Sholto Douglas, an engineer at Anthropic, took to X to announce that their latest model, Claude Mythos, had cracked a century-old puzzle in combinatorial geometry: the Erdős unit distance conjecture. Dubbed a "clever and concise proof," th
Amazon to Acquire Decart for $7B, Securing Key AI Chip Adapter
Per the Wall Street Journal, Amazon is negotiating to purchase AI startup Decart at an estimated $7 billion valuation. If no competing bids arise, the transaction could close shortly, underscoring major tech firms’ urgency to secure computing resourc











