Glossary
CodecFlow
Section titled “CodecFlow”CodecFlow is an open-source infrastructure stack for robotics and physical AI.
It connects the work needed to build, simulate, train, deploy, monitor, and improve robots.
SimArena
Section titled “SimArena”SimArena is the browser workspace inside CodecFlow.
Builders use it to import robot models, create scenes, run simulations, collect data, and start training workflows without setting up a local robotics stack.
Fabric
Section titled “Fabric”Fabric is CodecFlow’s compute layer.
It routes heavier jobs such as cloud simulation, training, and runtime workloads to external compute resources.
optr is the runtime direction for robot programs.
It is built around See, Think, Act workflows: perception, reasoning, and action.
See, Think, Act
Section titled “See, Think, Act”See, Think, Act is a simple way to describe a robot brain.
The robot observes the world, decides what to do, then sends actions to the robot body or simulator.
Robot model
Section titled “Robot model”A robot model is a digital description of a robot.
It can include links, joints, meshes, collision shapes, mass properties, sensors, and control interfaces.
A scene is the environment where a robot runs.
It can include floors, objects, obstacles, lighting, physics settings, and task conditions.
Simulation
Section titled “Simulation”Simulation means running a robot in a virtual environment before testing it in the real world.
Teams use simulation to test behavior, collect data, and catch failures earlier.
Browser simulation
Section titled “Browser simulation”Browser simulation means the simulation runs inside a web browser.
This is useful for quick setup, demos, light testing, and early workflow iteration.
Cloud simulation
Section titled “Cloud simulation”Cloud simulation runs heavier simulation jobs on remote compute.
Teams use it when browser simulation is not enough for scale, speed, or engine requirements.
Headless simulation
Section titled “Headless simulation”Headless simulation runs without a visual interface.
Teams use it for automated testing, batch runs, regression checks, and CI-style robotics workflows.
URDF stands for Unified Robot Description Format.
It is a common format for describing robot structure, including links, joints, meshes, and physical properties.
MuJoCo XML
Section titled “MuJoCo XML”MuJoCo XML is a model format used by the MuJoCo physics engine.
Robotics teams use it for simulation, control, and reinforcement learning workflows.
Isaac Sim
Section titled “Isaac Sim”Isaac Sim is a simulation platform from NVIDIA.
Teams use it for robotics simulation, synthetic data, and GPU-accelerated workflows.
Genesis
Section titled “Genesis”Genesis is a physics and robotics simulation platform.
It is often used for large-scale simulation, robot learning, and synthetic data workflows.
Rapier
Section titled “Rapier”Rapier is a physics engine that can run in the browser.
SimArena can use browser-friendly engines like Rapier for lighter simulation work.
PhysX is a physics engine used in simulation and game environments.
In robotics workflows, it can help simulate rigid bodies, collisions, and physical interactions.
LeRobot
Section titled “LeRobot”LeRobot is an open-source format and toolkit for robot learning datasets.
SimArena can export demonstration episodes in formats like LeRobot so teams can use collected data for training.
Demonstration
Section titled “Demonstration”A demonstration is a recorded example of a task being performed.
In robot learning, demonstrations can teach a model how a robot should behave.
Episode
Section titled “Episode”An episode is one recorded run of a task.
It usually includes observations, actions, rewards, timing, and whether the task succeeded.
Observation
Section titled “Observation”An observation is what the robot or agent receives from the environment.
It can include camera images, joint positions, object states, sensor readings, or simulation state.
Action
Section titled “Action”An action is the command sent by a policy or controller.
It can move a joint, set a target velocity, choose a gripper command, or trigger another robot behavior.
Policy
Section titled “Policy”A policy is the model or program that chooses actions for a robot.
It takes observations as input and returns actions as output.
Imitation learning
Section titled “Imitation learning”Imitation learning trains a policy from demonstrations.
Instead of learning only through trial and error, the model learns from examples of the task being done.
Reinforcement learning
Section titled “Reinforcement learning”Reinforcement learning trains a policy through rewards.
The robot tries actions, receives feedback, and learns behavior that improves the reward over time.
Evaluation
Section titled “Evaluation”Evaluation checks how well a robot policy performs.
Teams use evaluation runs to measure success rate, failure cases, robustness, and behavior changes between versions.
Dataset
Section titled “Dataset”A dataset is a collection of recorded robot data.
It can include demonstrations, simulation outputs, images, actions, rewards, and task metadata.
Synthetic data
Section titled “Synthetic data”Synthetic data is data generated from simulation instead of the real world.
Robotics teams use it to train and test models before collecting expensive real-world data.
Runtime
Section titled “Runtime”A runtime is the environment where a robot program runs.
For robotics, this can include perception, planning, control, hardware interfaces, monitoring, and cloud or edge execution.
Edge refers to compute that runs close to the robot, often on the robot itself.
Robots use edge compute when they need low-latency control or local operation.
Cloud refers to remote compute.
Robotics teams use cloud compute for heavier workloads such as training, large simulations, dataset processing, and evaluation.
An agent is the system that decides what to do in an environment.
In robotics, an agent can be a learned policy, a planner, or a program that combines perception, reasoning, and action.
Workflow
Section titled “Workflow”A workflow is a connected sequence of steps.
In CodecFlow, a workflow can go from robot import, to scene setup, to simulation, to data collection, to training, to evaluation.