# Embodied AI: Vocabulary Hub: [[Topics/Embodied AI/Embodied AI|Embodied AI]] · [[Topics/Embodied AI/Sources|Sources]] Plain definitions. Each Microduck example comes from the [microduck](https://github.com/pollen-robotics/microduck) or [microduck_rl](https://github.com/pollen-robotics/microduck_rl) READMEs or from [Remi Fabre's Oct 23 post](https://x.com/RemiFabreRobot/status/2113710447912218861). ◻ marks a general definition from background knowledge. - **Embodied AI / physical AI:** ◻ AI that senses and acts through a physical body (robots, cars) rather than only producing text or images. "Physical AI" is NVIDIA's marketing term for the same idea. - **Policy:** ◻ the trained neural network that maps what the robot senses (joint angles, IMU, commands) to motor targets, many times per second. Microduck runs its policies on a "50 Hz control loop driving fifteen servos." - **Reinforcement learning (RL):** ◻ training by trial and error with a reward signal (for example, stay upright and follow the commanded speed). Fabre: "use reinforcement learning when balance is part of the movement." - **PPO (Proximal Policy Optimization):** ◻ the most common RL algorithm for legged robots. microduck_rl is "built on mjlab (MuJoCo Warp) with PPO." - **Simulation / MuJoCo / mjlab:** ◻ MuJoCo is a physics simulator, now open source under Google DeepMind. mjlab is a GPU robot-learning framework on top of MuJoCo Warp. Thousands of virtual ducks fall over in parallel so the real one doesn't have to. - **Isaac Sim / Isaac Lab:** ◻ NVIDIA's simulator and robot-learning framework, the main rival to the MuJoCo stack. The Microduck training repo uses mjlab, not Isaac. - **Sim2real (sim-to-real transfer):** ◻ getting a policy trained in simulation to work on real hardware. The main tricks are domain randomization and accurate actuator models. - **Domain randomization:** ◻ randomly varying mass, friction, motor strength, latency and so on during training so the policy holds up against real-world differences. It is used in microduck_rl. - **Actuator model (BAM) / backlash:** a model of how the real servos respond, including gear slop ("backlash"). microduck_rl models backlash on 14 servo joints. - **ONNX:** ◻ a portable file format for trained networks. Microduck policies are "exported to ONNX" and run on the robot's Rockchip RK3566. - **Policy hot-swap:** "At deployment the runtime hot-swaps these policies (walk / recover / trick)." The robot has separate trained skills, not one general brain. - **Fall recovery:** a learned get-up skill. The microduck_rl task list includes "VelStand: Walking + fall recovery in one policy" and "StandUp." This is the real tech behind @Yumzlef's "no rigid scripts" claim. - **IMU (inertial measurement unit):** ◻ an accelerometer plus gyroscope that tells the robot which way is down and how fast it is tipping. It's the main input to a balance policy. - **Teleoperation:** a human sends commands remotely. Microduck's high-level control (where to walk, look, when to do a trick) came from a gamepad or app. At GTC, "one engineer piloting each Microduck." Low-level balance stays autonomous, which is why both "it's autonomous" and "it's remote-controlled" can sound true. - **Odometry:** ◻ estimating your own movement from internal sensors. Fabre: "the robot estimating its own movement from its own internal sensors." - **SLAM (simultaneous localization and mapping):** ◻ building a map while tracking your position on it. Pirrone has shown SLAM running onboard, per Fabre. - **Spiking neural network:** ◻ a niche brain-inspired network type that communicates in discrete spikes. @Yumzlef's "neurons firing" graphics borrow its look. Microduck's policies are ordinary PPO-trained networks exported to ONNX. Neither README mentions spiking networks ❌. - **JSON-RPC / WebRTC:** Microduck's control protocol and video stream. "Every client — the app, the console, the gamepad, your script — sends exactly the same calls." - **Open Duck Mini:** ◻ Antoine Pirrone's earlier open-source, 3D-printed duck biped (inspired by Disney's BDX droids), the hobby project that Microduck grew out of.