The Vidar embodied AI mannequin from ShengShu makes use of simulated worlds as an alternative of bodily coaching information. Supply: Adobe Inventory, Vectorhub by ice
ShengShu Expertise Co. yesterday launched its multi-view bodily AI coaching mannequin, Vidar — which stands for for “video diffusion for motion reasoning.” Utilizing Vidu’s capabilities in semantic and video understanding, Vidar makes use of a restricted set of bodily information to simulate a robotic’s decision-making in real-world environments, mentioned the corporate.
“Vidar provides a radically totally different method to coaching embodied AI fashions,” acknowledged ShengShu Expertise. “Simply as Tesla focuses on vision-based coaching and Waymo leans into lidar, the {industry} is exploring divergent paths to bodily AI.”
Based in March 2023, ShengShu Expertise specializes within the improvement of multimodal giant language fashions (LLMs). The Beijing-based firm mentioned it delivers mobility-as-a-service (MaaS) and software-as-a-service (SaaS) merchandise for smarter, quicker, and extra scalable content material creation.
With its flagship video-generation platform Vidu, ShengShu mentioned it has reached customers in additional than 200 nations and areas all over the world, spanning fields together with interactive leisure, promoting, movie, animation, cultural tourism, and extra.
Vidar simulated coaching to speed up robotic improvement
“Whereas some corporations prepare bodily AI by embedding fashions into real-world robots and accumulating information by means of the bodily interactions that their robots encounter, it’s a way that’s pricey, hardware-dependent, and tough to scale,” mentioned ShengShu Expertise. “Others depend on purely simulated coaching, however this typically lacks the variability and edge-case information wanted for real-world deployment.”
Vidar takes a unique method, the corporate claimed. It combines restricted bodily coaching information with generative video to make predictions and generate new hypothetical eventualities, making a multi-view simulation that includes lifelike coaching environments, all inside a digital house. This enables for extra sturdy, scalable coaching with out the time, value, or limitations of physical-world information assortment, defined ShengShu.
Constructed on prime of the Vidu generative video mannequin, Vidar can carry out dual-arm manipulation duties with multi-view video prediction and even reply to natural-language voice instructions after fine-tuning. The mannequin successfully serves as a digital mind for real-world motion, mentioned the corporate.
Utilizing Vidu’s generative video engine, Vidar generates large-scale simulations to scale back dependency on bodily information, whereas sustaining the complexity and richness wanted to coach real-world-capable AI brokers. ShengShu mentioned Vidar can extrapolate a generalized sequence of robotic actions and duties from solely 20 minutes of coaching information. The corporate asserted that’s between 1/80 and 1/1,200 of the info wanted to coach industry-leading fashions together with RDT and π0.5.
ShengShu mentioned Vidar’s core innovation lies in its modular two-stage studying structure. Not like conventional strategies that merge notion and management, Vidar decouples them into two distinct phases for better flexibility and scalability.
Within the upstream stage, large-scale basic video information and moderate-scale embodied video information are used to coach Vidu’s mannequin for perceptual understanding.
Within the second downstream stage, a task-agnostic mannequin known as AnyPos turns that visible understanding into actionable motor instructions for robots. This separation makes it considerably simpler and quicker to coach and deploy AI throughout several types of robots, whereas reducing prices and growing scalability.
Vidar is designed to scale back the quantity of coaching information wanted to coach AI fashions. Supply: ShengShu Expertise.
Vidar a framework for scalable embodied intelligence
Vidar follows a scalable coaching framework impressed by language and picture basis fashions of the previous decade of AI breakthroughs. ShengShu mentioned its three-tiered information pyramid, spanning large-scale generic video, embodied video information, and robot-specific examples, makes for a extra versatile system, decreasing conventional information bottleneck.
Constructed on the U-ViT structure, which explores the fusion of diffusion fashions and transformer architectures for a large assortment of multimodal technology duties, Vidar harnesses long-term temporal modeling and multi-angle video consistency to energy bodily grounded decision-making.
This design helps fast switch from simulation to real-world deployment, which ShengShu mentioned is crucial for robotics in dynamic environments. It additionally minimizes engineering complexity, in accordance with the corporate,
ShengShu mentioned Vidar can facilitate robotics adoption throughout a number of sectors. From dwelling assistants and eldercare to sensible manufacturing and medical robotics, the mannequin allows quick adaptation to new environments and multi-task eventualities, all with minimal information, it added.
Vidar creates an AI-native path for robotics improvement that’s environment friendly, scalable, and cost-effective, ShengShu claimed. By remodeling basic video into actionable robotic intelligence, the corporate mentioned its mannequin can bridge the hole between visible understanding and embodied company.
Vidar has a modular studying structure. Supply: ShengShu Expertise
ShengShu marks milestones in multimodal AI
Vidar builds on the fast momentum of the Vidu video basis mannequin, mentioned ShengShu. The corporate listed statistics since its debut:
- Vidu reached 1 million customers inside one month
- Surpassed 10 million customers in simply three months
- Generated over 100 million movies by Month 4
- Reference-to-video technology exceeded 100 million by Month 8
- Complete generated movies now prime 300 million
ShengShu continues to increase the frontiers of multimodal AI, Vidar represents the following frontier—bringing generalization, generativity, and embodiment into one unified system.
Editor’s be aware: RoboBusiness 2025, which will probably be on Oct. 15 and 16 in Santa Clara, Calif., will embody tracks on bodily AI and humanoid robots. Registration is now open.
