EmbodiedJev lets Jev select robot subgoals and motor channels while local code owns geometry, collision checks, execution, and physical verification.
In the Jev path, code computes geometry, candidate actions, and safety boundaries before the model selects a subgoal plus X, Y, Z, and gripper channels from bounded options. The system observes again after each step; local code owns motion conversion, workspace limits, collision checks, and success criteria.
The repository separates model selection, deterministic control, and physical verification while recording inputs, probabilities, trajectories, call counts, and latency. It is useful for studying Jev in local robot decisions, not as a replacement for motion control, emergency stops, or hardware safety.
README, policy code, hierarchical control, and experiment notes at pinned commit f08de2e were reviewed on 2026-09-26, but experiments were not reproduced here. Success, cost, and latency figures are author-reported simulation results; different protocols, inputs, and budgets are not directly comparable or predictive of hardware performance.
No. Jev receives structured state prepared by code and selects a subgoal plus X, Y, Z, and gripper channels from bounded options. Motion conversion, workspace limits, collision checks, and success criteria remain in the local control layer.
Study bounded robot-action selection separately from geometry and low-level control.
Material analysed：
policies.py checks the selected action, exact probability keys and numeric distribution before accepting a model decision. A syntactically valid label alone is insufficient.
Inspect the supporting source ↗The README records low-probability cases as uncertain in batch evaluation rather than silently switching them to the rule baseline. Otherwise success counts would mix different controllers.
Inspect the supporting source ↗Use it to inspect a controlled simulation comparison. Retain motion limits, collision checks and success evaluation outside the semantic policy; simulation findings do not establish hardware safety.
Compare policy output validation with the README's threshold and provider settings. When comparing runs, keep task seeds, observation format, call budget and uncertainty handling identical.