AutomotiveCloudWatch

Glossary · Robotics and Physical AI

Physical AI and Foundation Models

92 terms in the Robotics and Physical AI glossary, each with its one-line definition.

92 terms, A to Z

  • Action Chunking Having a robot model predict a short sequence of future movements at once so motion is smoother and small errors do not pile up
  • Action Expert Dedicated module inside a vision language action model that turns scene understanding into smooth continuous robot motion
  • Action Head Final module of a robot model that converts its understanding into motor commands such as joint angles or gripper positions
  • Action Tokenization Converting continuous robot movements into discrete tokens so a language style model can predict them like words
  • Affordance What an object lets an agent do with it, such as a handle to pull or a button to press, used to help robots handle new objects
  • Attention Mechanism Core transformer operation that lets each input element weigh how much to focus on every other element
  • Autoregressive Model Model that generates its output one piece at a time, each based on all the pieces produced so far
  • Beijing Innovation Center of Humanoid Robots Beijing center backed by MIIT, the city and robot makers that builds the Tiangong humanoid, open datasets and a shared embodied AI platform
  • Code as Policies Approach where a large language model writes short programs that call robot perception and motion functions instead of driving joints
  • Contrastive Language-Image Pretraining CLIP OpenAI method that trains a model to link images with matching text, letting robots recognize objects from plain descriptions
  • Cosmos NVIDIA family of world foundation models that generate and reason about physical world video for robot and AV training data
  • Cross-Embodiment Learning Training one robot control model on data from many robot types so skills learned on one body help others
  • Denoising Diffusion Probabilistic Model DDPM Specific diffusion model recipe that generates outputs by cleaning random noise step by step, used in early diffusion robot policies
  • Denoising Steps Number of refinement rounds a diffusion or flow model runs to turn noise into an action, trading quality against reaction speed
  • Diffusion Model Generative AI that creates outputs by reversing a noising process, now widely used to generate robot movements
  • Diffusion Transformer DiT Diffusion model that uses a transformer as its denoising network, used for world video generation and robot actions
  • DINOv2 Meta family of self supervised vision models whose features capture geometry well, used as robot vision encoders
  • Embedding List of numbers a model uses to represent a word, image patch or sensor reading so similar things get similar values
  • Embodied AI Field of AI where an agent with a body senses and acts in a real or simulated world, covering robots and virtual agents
  • Embodied Chain-of-Thought Reasoning ECoT Training a robot model to write out a plan and object locations before acting, improving generalization at some speed cost
  • Embodied Reasoning ER Reasoning about objects, spatial relations, grasp points and task steps in the way a robot needs in order to act
  • Embodied Reasoning Question Answering ERQA Google DeepMind benchmark of multiple choice image questions testing whether a model understands the physical world well enough for robots
  • Embodiment Particular physical body of a robot, including its shape, joints, sensors and motors
  • Emergent Capability Skill that appears in a large AI model without being trained for directly, often only once the model is big enough
  • FAST Action Tokenizer FAST Physical Intelligence method that compresses robot motion into few tokens so language style models can drive fast dexterous control
  • Flow Matching Generative technique related to diffusion that moves from noise to output along direct paths, usually needing fewer steps
  • Gato Early DeepMind generalist agent that used one model to play games, caption images, chat and stack blocks with a robot arm
  • Gemini Robotics Google DeepMind family of models that bring Gemini into the physical world for direct robot control and embodied reasoning
  • GEN-0 Generalist AI family of robot foundation models trained on large real world manipulation data and pitched as evidence of robot scaling laws
  • Generalist Policy Single robot control model trained to do many tasks in many settings, often across several robot types
  • Generative AI Accelerator Challenge GENIAC METI and NEDO program giving Japanese developers subsidized compute and support to build their own generative AI foundation models
  • Genie 3 Google DeepMind world model that generates interactive 3D worlds from text in real time for people or AI agents to explore
  • Genie Operator-1 GO-1 AgiBot robot foundation model trained largely on its AgiBot World dataset, an early prominent Chinese VLA style model
  • GR00T N1 NVIDIA openly released humanoid robot foundation model that developers adapt to their own robots
  • Grounding Linking words and ideas to the specific real objects and actions a robot can see and perform
  • Helix Figure AI model that takes camera images and instructions and directly controls its humanoid robots end to end
  • Hi Robot Physical Intelligence system where a high level model turns open ended requests and feedback into simple commands for a low level policy
  • Hierarchical Policy Layered robot control design where a high level picks the next step or skill and a low level moves the motors
  • Huisi Kaiwu Beijing humanoid innovation center platform meant to let one AI system control many robot bodies for sorting, assembly and packing
  • In-Context Learning ICL Model ability to pick up a new task from examples or instructions in its input at run time without changing its weights
  • Instruction Following How well a robot or AI model does what it is told, including commands phrased or combined in new ways
  • Isaac GR00T NVIDIA humanoid development program bundling AI models, simulation and data tools and onboard robot computers
  • Language Conditioning Feeding a written or spoken instruction into a robot model so one model does different jobs depending on what it is told
  • Large Behavior Model LBM Toyota Research Institute name for large multitask robot policies learned from many demonstrations and guided by language
  • Large Language Model LLM AI model trained on huge amounts of text to predict the next word, now used by robots as a planning brain
  • Latent Action Pretraining LAPA Method that learns compact action codes from unlabeled videos and later maps them to real motor commands with a little robot data
  • Latent Space Internal compressed space where a neural network represents what it learned, summarizing inputs rather than storing them raw
  • Mixture of Experts MoE Network design that routes each input to a few specialized subnetworks so a model can be large without using all of it every time
  • Moonshot Goal 3 Japanese government research goal for self learning AI robots that act alongside people by 2050
  • Multimodal Model AI model that works with more than one kind of data, such as images, text, audio, touch or joint readings
  • National and Local Co-Built Humanoid Robotics Innovation Center Shanghai humanoid center jointly funded by China's national and city governments, developer of Qinglong and OpenLoong
  • NVIDIA Three-Computer Model NVIDIA framing of physical AI as three computers, one to train, one to simulate and one inside the robot to run models
  • Octo Open source generalist robot policy from university researchers that anyone can download and fine tune for their robot
  • Open-Weight Model AI model whose trained parameters are published for anyone to download and run on their own hardware
  • OpenLoong Open source humanoid project from Shanghai's humanoid innovation center that published its Qinglong robot design and software
  • openpi Physical Intelligence public code repository for downloading, running and fine tuning its pi0 family of robot models
  • OpenVLA Openly released 7 billion parameter vision language action model that became the standard research baseline
  • OpenVLA-OFT Improved fine tuning recipe for OpenVLA that outputs whole action chunks at once for faster, more accurate control
  • PaliGemma Google open vision language model that became a common backbone for robot control models such as pi0
  • PaLM-E Early Google model combining a large language model with robot camera and sensor inputs for planning and answering questions
  • Physical AI AI that senses, understands and acts in the real world through machines such as robots and self driving cars
  • pi*0.6 Physical Intelligence robot model that improves from its own on the job experience and human corrections using reinforcement learning
  • pi0 Physical Intelligence robot foundation model that reads images and instructions and directly outputs smooth robot motion
  • pi0.5 Physical Intelligence follow on to pi0 built to generalize to unseen places such as unfamiliar homes
  • Post-Training Training applied after a model's large initial run to specialize it or make it more reliable at particular tasks
  • Pretraining First large training stage where a model learns general knowledge from broad data before being adapted to specific jobs
  • Qwen-VL Alibaba family of open vision language models, widely used as backbones for open VLAs, especially in China
  • Real-Time Chunking RTC Technique that keeps a robot executing its current action chunk while the model computes the next, then blends them smoothly
  • RoboBrain BAAI open embodied reasoning model that plans tasks, finds grasp points and sketches paths for robots from images and instructions
  • RoboCat Google DeepMind robot agent that learned new tasks from few examples and improved by generating its own training data
  • Robot Foundation Model RFM Large AI model trained on varied robot and internet data so it can adapt to many tasks, places and robots
  • Robotics Diffusion Transformer RDT-1B Tsinghua University open bimanual robot model that generates motion with diffusion and pools data across two armed robots
  • RT-1 Early Google robot model that turned camera images and text instructions into actions, trained on 130,000 real demos
  • RT-2 Google DeepMind model that first showed a web trained AI model could directly control a robot, coining vision language action
  • RT-X Google DeepMind and partner models trained on the pooled Open X-Embodiment dataset across many robot types
  • SayCan Early method combining a language model's view of sensible steps with a robot's estimate of which steps it can actually do
  • Scaling Laws Patterns showing how model performance improves predictably with more data, larger models and more compute
  • SigLIP Google image and text model similar to CLIP but trained with a sigmoid loss, popular as the vision part of robot models
  • SmolVLA Hugging Face small open VLA model that runs on a gaming GPU and controls low cost hobby robot arms
  • Spatial Reasoning Understanding where things are in 3D and how they relate, such as left of, on top of or within reach
  • System 1 / System 2 Architecture Robot AI design pairing a slow reasoning model with a fast motor control policy
  • Transformer Neural network design behind modern AI that processes a sequence all at once with every part attending to every other
  • UniSim Learned video simulator that predicts how a scene looks after an action, letting robots practice inside it
  • V-JEPA 2 Meta world model that learned physical dynamics from over a million hours of video and can plan simple robot actions
  • Video Generation Model AI model that creates realistic video from text or images, reused in robotics as imagined worlds or training data sources
  • Vision Encoder Part of an AI model that turns camera images into numbers the rest of the model can work with
  • Vision Transformer ViT Transformer applied to images by splitting each picture into small tiles treated like words
  • Vision-Language Model VLM AI model that understands images and text together, the usual backbone of robot control models
  • Vision-Language-Action Model VLA AI model that reads camera images and a plain language instruction and directly outputs robot movements
  • VoxPoser Stanford method where a language model writes code that builds 3D value maps of where a gripper should go for a planner to follow
  • World Foundation Model WFM Large model that learned how the physical world looks and behaves from video, then tailored to a robot, vehicle or setting
  • World Model AI model that predicts how the world will change after an action so a robot can imagine outcomes before acting

Full entry for members

Every term has a full entry for members: the plain-words definition, the practitioner note and the typed links between terms.

See membership

ACW Robotics and Physical AI Glossary, edition 2026.11.1. The headword and one line are open to everyone.