JackAILab's picture
Add files using upload-large-folder tool
6aa6c12 verified
|
Raw History Blame Contribute Delete
2.36 kB

Shared infrastructure diagnostics

This package contains harness-owned fixtures, not participant model policies. constant.py returns a supplied fixed action chunk through the same Policy interface used by real methods. It performs no model inference and does not plan actions from observations. Its descriptor labels every run infrastructure_validation_only.

The CLI name remains --policy constant; action validation, chunk length and execution budgets are unchanged. Using this policy with a native adapter can move the robot or simulator, so only run it as an explicitly authorized diagnostic with valid actions and environment safety checks. It is not a task capability score or a named-model baseline.

Shared diagnostics are included in new harness_source/ snapshots. Old frozen campaigns and their recorded source inventories retain their original bytes; there is no AgentPolicies.diagnostics compatibility shim in the current source.

robot_state_hold.py is an explicit one-control SIM-only sensor/target wiring check. The operator must first confirm that the identically named public sensor and action fields mean absolute robot targets in the same units and frame. It echoes that state once through the ordinary authenticated SIM server client, not through a private simulator handle. The server must declare diagnostic=true, and --confirm-absolute-state-targets is mandatory. It does not infer a task action or guarantee physical holding, and must never be used on real hardware. Its one fixture decision has zero model requests and is not a capability score. Keep these artifacts separate from scored model attempts.

Native asset, rendering, reset and control readiness tools belong under the corresponding benchmark's benchmarks/.../environment/. Participant policy behavior belongs under AgentPolicies/<profile>/.

CPU-only checks for this package:

python -B -m unittest benchkit.checks.test_diagnostics -v

simpler_mailbox.py retains the fixed seven-zero-action mailbox-v2 diagnostic formerly under SimplerEnv's environment directory. It is an explicitly selected historical transport fixture, not a model or a general robot-safe action. It requires matching v2 / decision-budget settings; it does not support the current v3 mailbox by implication. Do not use it on real hardware or count it as a score.