波士顿动力x TRI联手！使用大型行为模型（LBM）训练Atlas！目标“AI通才机器人”

Core Viewpoint - The development of humanoid robots requires them to perform multiple tasks, including manipulating various objects and maintaining balance in unexpected situations. Large Behavior Models (LBM) are crucial for cultivating these core capabilities in humanoid robots [1][2]. Group 1: Collaboration and Development - Boston Dynamics has partnered with Toyota Research Institute (TRI) to develop LBM for the Atlas humanoid robot, utilizing end-to-end language modulation strategies to assist in long-term manipulation tasks [2]. - The strategy consists of four processes: collecting behavior data through remote operation, processing and annotating the data, training neural network strategies, and evaluating these strategies with a testing suite [3]. Group 2: Strategy Principles - Boston Dynamics follows three core principles in strategy formulation: maximizing task coverage through a remote operation system, adopting a multi-task training policy for better generalization, and building a robust infrastructure for rapid iteration and scientific rigor [5][9]. Group 3: Hardware and Software Configuration - The Atlas robot has 78 degrees of freedom (DoF) for extensive movement and flexibility, while the Atlas MTS focuses on pure manipulation tasks with 29 DoF [9]. - The remote operation system uses HDR stereo cameras for situational awareness and integrates with a model predictive controller (MPC) to ensure precise operations while maintaining balance [9][10]. Group 4: LBM Technology and Simulation - The LBM architecture is based on TRI's LBM, utilizing a diffusion transformer with 450 million parameters to predict actions based on sensory input and language prompts [11]. - Simulation technology plays a key role in development, allowing for efficient training and evaluation while sharing data pipelines and reducing costs [11]. Group 5: Enhanced Capabilities - Through LBM training, Atlas has surpassed traditional robotic limitations, enabling it to autonomously respond to unexpected situations and perform complex long-range tasks [12][14]. - The robot can execute a variety of tasks, from simple pick-and-place actions to complex operations like manipulating a 22-pound (9.9 kg) tire, showcasing the advantages of "learning by demonstration" [16]. Group 6: Future Directions - Boston Dynamics aims to expand its "data flywheel" to improve throughput, quality, task diversity, and difficulty while exploring new algorithmic concepts [19].