Physical AI Plus

2026 is shaping up to be the Year of data collection for embodied AI. As more embodied AI companies validate the scaling laws of egocentric data, a new question emerges: How can we collect massive amounts of high-quality, useful data faster and at lower cost? Maniformer’s answer is MEgo View + MEgo Engine — an integrated solution designed to address real pain points across data collectors, data collection centers, and model companies, connecting the full egocentric-data pipeline from hardware-based data capture and data processing to model training. #PhysicalAI #DataCollection #EgocentricData #Maniformer #EmbodiedAI #Robotics #AgiBot

Dyna Robotics is making a major move! Trained on 1 million hours of human first-person (egocentric) data, the WAM model Dyna-2 demonstrates that human manipulation videos follow a scaling law — and it outperforms VLA models. Now the big question is: how much egocentric data will it take to achieve the “GPT moment” for embodied AI models? #PhysicalAI #Robotics #EmbodiedAI #Dyna #DataCollection #EgocentricData #Data #FoundationModels #GPT #EmbodiedModels

After VLA, world models are becoming the next major battleground for embodied intelligence. KAI from Kinetix AI is exploring a different path: combining a physical world model, first-person data, and highly human-like robotic embodiments to help robots truly understand the real world. #PhysicalAI #EmbodiedAI #Robotics #HumanoidRobots #KAI #Kinetix #worldmodel #WAM #VLA

With the shortage of embodied AI data, which type of data is more valuable: teleoperation data, simulation data, ego-centric data, or human motion data? Professor Su Hang, Research Associate Professor at the Department of Computer Science and Technology at Tsinghua University, believes that all types of data are important. The key is to leverage different data sources together to enhance the pre-training capabilities of embodied AI models. #EmbodiedAI #Data #AI #FoundationModels #PreTraining #Robotics #PhysicalAI #Maniformer

An impressive real-world demonstration of a World Action Model is here! This video shows the actual deployment of GigaAI’s WAM model, with no speed acceleration. #EmbodiedAI #FoundationModels #PhysicalAI #WorldModel #WorldActionModel #WAM #VLA

Do robots really need five-fingered dexterous hands? When answering this question, Tony Zhao, Co-founder and CEO of Sunday Robotics, highlighted the 80/20 trade-off between cost and performance. #PhysicalAI #Robotics #EmbodiedAI #DexterousHands #DexterousManipulation

Which is better, VLA or WAM? Tencent Robotics X Lab’s answer is not to commit to a single model architecture. Instead, it starts from the perspective of building a complete embodied agent. Its Apexio system adopts a hierarchical architecture. #PhysicalAI #EmbodiedAI #Robotics #Tencent #RoboticsX #Apexio #EndToEnd #IntelligentAgent #AIagent #VLA #WAM

Dyna Robotics is moving away from the VLA approach and choosing WAM—World Action Models—instead. How does Dr. Ma, Co-founder and Chief Scientist of Dyna Robotics, explain this decision? #Dyna #DynaRobotics #embodied #physicalai #Robotis #robot #model #worldactionmodel #VLA

Which is more powerful, VLA or WAM? Let’s hear how Prof. Danfei Xu, Assistant Professor at the School of Interactive Computing at Georgia Tech and NVIDIA AI Researcher, answers this question. Ultimately, it is still about being goal-oriented — the key is to evaluate which approach can truly improve a robot’s capabilities. #embodiedai #physicalai #robotics #NVIDIA

The model architecture for Physical AI has yet to converge. Some say VLA is dead—but what does a research scientist at Physical Intelligence think? #physicalAI #physicalintelligence #PI #π #VLA #worldmodel #worldactionmodel #π07