# GPT-6 Astra / Codex-in-the-Loop RoboDojo Pilot-20 v1

Registered before pilot trials on 2026-09-08. Scope: 20 tasks, seeds 0/1/2, one first public layout per seed, 60 trials. These are 60 evaluations, not the official 6300-episode benchmark. Native seed manager chooses the layout from each seed directory. No scene-layout JSON contents are read by the controller.

## Selection

Four tasks per capability dimension, fixed before observing pilot results:

| Dimension | Task | Reason |
| --- | --- | --- |
| Generalization | push_T | planar pushing and orientation |
| Generalization | stack_bowls | grasping, nesting and stacking |
| Generalization | fold_clothes | deformable object manipulation |
| Generalization | pour_liquid_into_cup | container manipulation and liquid |
| Precision | deposit_coin | small object and narrow opening |
| Precision | insert_key | alignment and insertion |
| Precision | insert_tubes | peg/tube placement |
| Precision | play_Xylophone | precise repeated contact |
| Long-Horizon | fill_egg_holder | repeated placement into receptacles |
| Long-Horizon | fill_pen_holder | sequential slender-object handling |
| Long-Horizon | put_bottles_into_dustbin | repeated grasp, transport and release |
| Long-Horizon | play_stacking_toy | ordered assembly |
| Memory | press_by_number | visual digits, counting and confirmation |
| Memory | swap_T | preserve identity and swap planar objects |
| Memory | swap_blocks | preserve identity and swap grasped objects |
| Memory | cover_blocks | remember targets and cover them |
| Open | general_pickup | instruction-grounded target selection |
| Open | align_blocks | relational spatial arrangement |
| Open | solve_equation | symbolic instruction and manipulation |
| Open | stack_blocks_by_language | language-grounded stacking |

## Controller and information boundary

One interactive GPT-6 Astra session in xhigh mode controls all trials through the historical Codex-in-the-Loop bridge, using a fixed library of numerical motion primitives and CuRobo inverse kinematics. The model/mode provenance is confirmed by the experiment operator; it was not serialized into each individual trial artifact. No separate paid model API is started. The policy may receive only RGB images from the standard head/wrist cameras, the standard language instruction, and left/right arm joint positions, gripper positions and end-effector poses. Public robot geometry and static camera calibration may be used. No depth/segmentation, privileged object pose, object identity labels, scene-layout answers, reward internals, future frames, reference trajectories or expert policy actions are supplied to the controller.

The controller stores a concise visible decision rationale with every action command. It can adapt actions from current observations, but task/reward code is not edited. The evaluator may access scene state to run physics and calculate scores; that state is not exported to the controller. Native termination and native task step limits remain unchanged. Giving up executes a stationary action through the remaining native step budget and is retained as a policy failure if the native evaluator says so.

Prior knowledge: a previous development experiment solved press_by_number seed 0 layout 0 and inspected its public task source/configuration. This task is a regression/control task and is not claimed as previously unseen or zero-shot. Other tasks are new to this controller. The experiment is an adaptive tool-using agent study, not a frozen neural-policy reproduction. Any infrastructure/controller-code repairs during the campaign must be logged with affected trials; policy failures are not silently rerun.

## Logging and comparisons

Preserve task/seed/layout identifiers, software/code hashes, instructions, decision-point camera images, robot observations, action arrays, concise decision rationales, timestamps, native result JSON, all three native videos, and process logs. Report runtime, number of decision chunks and executed/action counts where available. API token spend is unavailable in this conversation; never substitute estimates for measured usage.

Report each task as 0/3, 1/3, 2/3 or 3/3 successful, mean native score, and individual trial results. Show raw counts alongside percentages and Wilson 95% intervals. Record infrastructure failures separately from policy failures, with both coverage and all-planned-trial completion visible.

Use the official website snapshot dated 2026-09-05, downloaded 2026-09-08. Compare only the same selected tasks against existing per-task scores; do not run baseline policies. Missing baseline task values remain missing. Equal four-task dimension selection permits a balanced subset mean, but this is neither an official overall score nor an official rank. Task sampling, sample count, model execution style, temporal budget and prior development differ from the leaderboard. No official submission is made.

All original pilot results, including failures, remain in the record. Any later improved controller belongs to a new experiment version.