Keep nearby
Robotics vocabulary.
Use these terms consistently when reading the code and discussing results.
- Simulation model
- The world description: bodies, joints, physical parameters, cameras and actuators.
- Simulation state
- The changing quantities at an instant, such as joint positions and velocities.
- Observation
- The information deliberately exposed to an operating component. It is usually a subset of simulator truth.
- Action
- A control request. Our preparation pendulum uses torque; the future SO-101 interface uses documented joint targets.
- Joint / actuator
- A joint constrains relative motion; an actuator applies a controllable input associated with that motion.
- Control rate / physics rate
- How often action requests update / how often the simulator advances its dynamics.
- Coordinate frame
- An origin and axis directions used to describe a position or orientation.
- Forward / inverse kinematics
- Joint configuration to endpoint pose / desired endpoint pose to possible joint configurations.
- Policy
- A rule or learned model that maps observations to actions.
- ACT
- Action Chunking with Transformers, an imitation-learning policy approach that predicts sequences of actions.
- VLM / VLA
- Vision-language model / vision-language-action model. A VLM text response alone is not a validated motor action.
- Teacher / evaluator
- A demonstration-generating reference / an independent component judging what happened.
- Seed / randomization
- A reproducible random-number initialization / deliberately varied scene conditions.
- Ground-truth leakage
- Giving the deployed policy hidden simulator answers that the declared observation interface does not permit.
- MPS / OpenVINO
- PyTorch's Apple Metal execution backend / an inference optimization and runtime toolkit used for the Intel target.