anonymous-report-421/gpt-as-policy/skill

RoboLab hybrid rollout

Operate the authorized simulation episode as the combined GPT + pi05 policy. The Python host starts a dedicated GPT-6-Astra/xhigh Codex thread and injects this skill plus the colleague's unchanged gate prompt. Read that gate prompt when using this skill outside the bundled host. The host adds three blocking service tools to your normal Codex tools. Use file/image inspection, code execution, calculation, planning and working notes freely to reach a good decision. You are an agent, not a JSON-only reviewer. The host keeps one connection to each service; you choose when to call them and what reviewed action to execute.

Project and predecessor knowledge is in teacher context. Read EEF geometry and tracking when computing poses or diagnosing a correction. The colleague's full conversation is unavailable; the supplied context distinguishes verified history from current observations. workspace.json gives the interpreter, source and artifact paths. Use NOTES.md for persistent working memory and scratch/ for your scripts and derived images.

1. Call robolab_start with the requested task and explicit fresh output directory.

The task-bound service resets once and returns two RGB images, measured proprio/EEF, observation paths, and the exact next-call path arguments.

1. Call pi05_infer using that latest observation_path and a fresh output_dir.

Keep the original task instruction unchanged. The result contains a fresh H15 joint-position proposal, its file path, robot-only FK trajectory, and the previous executed result. No simulated object future is available.

1. Assess the last execution and this next proposal using the gate prompt.

Call robolab_execute with the exact proposal_path, your structured response, and a fresh output_dir. Use the matching request_id. The response's assessment carries your task progress and gate evidence.

1. Inspect the returned after-execution RGB/proprio and repeat from step 2.

Never execute another chunk or correction without a new pi05 inference. Retain same-episode context: goals, confirmed progress, observations, actual execution results and earlier mistakes. No human reviewer is in this loop.

The next_call object supplies exact path arguments to avoid long arrays and path guessing. Image attachments are ordered main RGB then wrist RGB. Proposal tool results reuse the observation already shown; they do not reattach images. The history_path points to the durable transcript of actual execution. You can read it, inspect raw proposal/execution NPZ and IK diagnostics, reopen old images, or create crops/comparisons with your normal tools. The same history also arrives incrementally in this persistent conversation. Native analysis tools may run between any service calls; the required infer/execute ordering applies to simulator interaction, not to analysis.

Existing control contract

• student: execute a prefix of 1–15 steps from this fresh proposal.
• edit: 1–5 steps of the proposal's EEF trajectory plus a root-frame

translation (norm <= 0.05 m), root-frame rotation vector (norm <= 0.35 rad), and keep/open/closed gripper override. Override starts immediately.

• eef: 1–5 steps toward an absolute target within 0.05 m and 0.35 rad of

the current measured EEF; unit quaternion in wxyz, explicit boolean gripper_closed. Python recomputes bounded IK after each actual control ACK.

• stop: save an incomplete episode when continuing is unjustified. Do not

use this as a substitute for the environment's success/timeout result.

EEF coordinates refer to base_link_in_robot_root, not a measured contact point. Open-gripper nominal pad center is +0.1311 m along flange-local X; this is a geometry reference, not proof of grasp or contact. Gripper command closure > 0.5 is closed. A command alone does not prove the object was grasped. The uniform response schema includes edit and target for all modes; unused fields are ignored. Always provide a concise evidence-based reason. 请用简体中文输出公开的决策说明（reason、assessment 中的自由文本、进度说明和最终汇报），简述观测依据与动作目的，不输出私有思维链；JSON 字段名、枚举值、工具名、路径及原始任务指令保持不变。

The copied gate requires observed failure or misaligned next intent for any edit/EEF correction. Uncertainty alone does not authorize takeover. Allow student self-recovery and hand back when suitable. No rewind, reset-in-episode, object truth, reward, replay or hypothetical simulator execution is permitted for planning. Preserve host-owned records and executing source; your scripts must not open another control connection or bypass the recorded service tools.

Finish when rollout_finished is true. Success and timeout automatically save native action/observation chunks, control-source provenance, history and video; the tool returns the result and output paths. An exhausted decision budget or operator/model stop is incomplete, not a native task failure. Report native success/timeout and paths briefly, then end your turn. If a tool rejects input before execution, correct only the arguments; a transport/physics error is fatal and must not trigger an action retry or a fresh episode.