{"id":"embodied-policy-evidence","name":"Learned action policies need embodiment evidence","shortName":"Learned action policies need embodiment evidence","role":"Observation/action encoding, training trajectory distribution, policy rate and downstream control constraints.","description":"A learned robot policy maps observations into an action representation trained for a particular embodiment and task distribution. An action token is not self-describing: a codec must map it to units, coordinate conventions, action dimensions and an update interval. A policy may produce an in-range token whose decoded action is unsuitable for the current geometry or dynamics. RT-2 gives primary context for learning action-token outputs from robot trajectories and vision-language data. The five-bin codec below is an authored example, not RT-2’s actual encoding or a running VLA model. Its purpose is to connect a symbolic output to a bounded numeric command and to keep that command distinct from a successfully completed physical task.","mechanism":["The observation contract specifies sensor inputs, timestamps, frames and any task-language context.","A learned or recorded policy produces an action representation associated with a specific model and embodiment.","The action codec maps symbols to declared command units, bounds and timing.","Runtime constraints and feedback evaluate feasibility and observed completion; token validity alone is insufficient."],"learningObjective":"Decode a bounded action representation into physical units and distinguish valid output from feasible motion.","engineeringExample":"An authored one-axis codec maps integer bins 0 through 4 uniformly onto velocities −0.2, −0.1, 0, 0.1 and 0.2 m/s. Bin 4 therefore commands +0.2 m/s; under an ideal constant-velocity interval of 0.2 s it requests 0.04 m displacement. Bin 5 is outside the schema and must be rejected. A valid bin 4 may still point into an obstacle, so range validation cannot replace geometry or feedback. These numbers illustrate action semantics without claiming that a language model produced or successfully executed them.","misconception":"A legal action token is not proof that the robot can safely or successfully execute it.","check":{"question":"Under the stated codec, what displacement does bin 4 request over 0.2 s?","options":["0.04 m","0.4 m","Four meters because the token is 4"],"answer":0,"explanation":"Bin 4 maps to 0.2 m/s; multiplying by 0.2 s gives 0.04 m under the ideal velocity assumption."},"checks":[],"prerequisites":["robot-goal-plan-action","robot-feedback-control"],"related":["tool-loop","evaluation"],"specs":[{"label":"Mechanism boundary","value":"What makes an action token meaningful to a robot?","context":"Teaching codec only; not a published model’s tokenization. Actual motion requires compatible control, geometry and dynamics; a requested displacement is not observed completion."},{"label":"Evidence type","value":"Authored worked example","context":"Worked explanation with a bounded numeric or state example. This lesson does not claim an executing implementation of an external model or platform."}],"interfaces":[{"name":"Declared input and state","direction":"in","signal":"The observation contract specifies sensor inputs, timestamps, frames and any task-language context."},{"name":"Checked outcome","direction":"out","signal":"Decode a bounded action representation into physical units and distinguish valid output from feasible motion."}],"tradeoff":"Discretized actions simplify an output vocabulary but trade resolution against representation size and still depend on the trained embodiment.","whyAI":"Vision-language-action systems connect learned representations to actuation through codecs, constraints and feedback.","sourceIds":["automation-rt2","automation-robot-control"],"sources":[{"title":"RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control","url":"https://arxiv.org/abs/2307.15818"},{"title":"Lynch and Park · Modern Robotics 11.3","url":"https://modernrobotics.northwestern.edu/nu-gm-book-resource/11-3-motion-control-with-velocity-inputs-part-1-of-3/"}],"science":[{"title":"Learned action policies need embodiment evidence · explicit relationship","equation":"This authored uniform codec maps an integer bin to velocity, then integrates that ideal velocity over one interval.","explanation":"An authored one-axis codec maps integer bins 0 through 4 uniformly onto velocities −0.2, −0.1, 0, 0.1 and 0.2 m/s. Bin 4 therefore commands +0.2 m/s; under an ideal constant-velocity interval of 0.2 s it requests 0.04 m displacement. Bin 5 is outside the schema and must be rejected. A valid bin 4 may still point into an obstacle, so range validation cannot replace geometry or feedback. These numbers illustrate action semantics without claiming that a language model produced or successfully executed them.","assumptions":"Teaching codec only; not a published model’s tokenization. Actual motion requires compatible control, geometry and dynamics; a requested displacement is not observed completion."}],"evidenceNotes":"This is a worked explanation, not an executing external platform, model or robot. The quantities and fixtures are authored. Source scope is recorded in the automation learning review manifest; no source benchmark or private system behavior is claimed. Teaching codec only; not a published model’s tokenization. Actual motion requires compatible control, geometry and dynamics; a requested displacement is not observed completion.","level":"model","branch":"Observations become physical actions","subbranch":"Learned action policies need embodiment evidence","scale":"Logical state, task evidence and declared units; no physical assembly scale","kind":"system","workstream":"W11","visualFamily":"software-flow","labId":"library","labScope":"The worked example explains this mechanism. General numerical labs are a reference; they do not execute this lesson’s full mechanism.","history":[],"workedTrace":{"bins":5,"maxSpeedMS":0.2,"selectedBin":4,"intervalS":0.2,"velocityMS":0.2,"displacementM":0.04}}