After seeing so many GPT-6 Astra robot demos on X, we had to try it on RoboDojo — a benchmark that is much less forgiving than a one-off demo.
We ran 20 RoboDojo tasks, three seeds each. Astra got three camera views, the task instruction and the robot’s current state. It decided where to move each arm, how to orient the wrist and when to open or close the gripper; the usual robot stack handled the low-level motion.
GPT-6 Astra · xhigh · 20 tasks × 3 seeds · successes and failures both included.
This was a small pilot: 20 selected tasks, one public layout per seed, and a shared adaptive conversation across runs. Removing the button-counting task, which had prior development, changes the result to 8/57 (14.0%). It is not an official RoboDojo submission.
Astra makes a decision every chunk; ordinary robotics software handles the motion in between.
At each decision point, Astra gets three RGB camera views, the task instruction and the robot’s current state. It chooses which arm to use, where the end effector should go, how the wrist should be oriented, how far to close the gripper and how long that move should last.
A small controller turns those choices into smooth trajectories. CuRobo and the robot interface handle inverse kinematics and execution. When the chunk finishes, Astra sees the scene again and decides whether to continue, correct the move or try something else.
Read the protocol ↗Head + wrist RGB, instruction, joint states, grippers and end-effector poses.
Which arm, where to move, wrist orientation, gripper setting, duration, and a short note about why.
move, grip, hold or home; interpolation and inverse kinematics fill in the motion between decisions.
The next image decides whether the plan continues, gets corrected, or is abandoned.
Astra received camera images, robot state and public static camera/robot calibration. It did not receive simulator object positions, depth, segmentation, reward state, reference actions or expert trajectories. We inspected the bridge code; the full interactive session did not undergo an independent information-isolation audit. The button-counting task also had prior development and is reported separately.
Astra first tried to lift the ruler. The next camera view showed it was still flat on the table, so the plan changed: close the gripper, push the ruler from behind, and use its straight edge to line up the blocks.
After checking the result, the arms moved back toward their starting posture. All three seeds passed the RoboDojo evaluator.
Public camera calibration turns a chosen image point into a ray. The controller intersects that ray with an estimated contact-height plane, then chooses an end-effector orientation and gripper state. The height comes from the visual estimate; the simulator does not hand over object depth.
See the actual recovery commands ↗Choose a task and a seed to see the RoboDojo score and the decision notes saved for that run. The full 180-video archive stays out of GitHub so this page remains quick to load.
| Task | Seed 0 | Seed 1 | Seed 2 |
|---|
No tasks match these filters.
Each run keeps its RoboDojo score, short decision notes, primitive counts and submitted action counts. The full RGB/action archive stays in the local research workspace to keep this public page lightweight.
These are the short notes saved with each action command, not internal chain-of-thought traces.
There are only four tasks in each capability category, so these cards are best read as clues about failure patterns rather than category-wide estimates.
Three runs show where the system succeeded, where it recovered, and where a sensible plan still fell apart in the physics.
ALIGNMENT · 3/3The first lifting attempt did not work. Astra switched to pushing the ruler with a closed gripper and used its straight edge to line up the blocks. All three seeds passed.
LANGUAGE STACKING · 2/3Two seeds recovered from placement errors and completed the requested stack. In the third, the requested order was tracked but the farther orange block was never acquired reliably.
EQUATION · 0/3The recorded rationales identified plausible missing symbols. Grasping and placement still failed, so all three scores stayed at zero.
Published task-level means over the same 20 task names give 18.3% a useful scale. The rollout protocol is different, so this comparison is context rather than a leaderboard rank.
| System | Success rate | Mean score / 100 |
|---|
Source: RoboDojo public leaderboard, snapshot dated September 5, 2026 and downloaded September 8. Baseline policies were not rerun, and their rollout conditions differ from this pilot. Use the table as context, not as an official ranking.
Each clip puts all three seeds side by side: two tasks with successful runs, one symbolic failure, and one precision task that made partial progress.
Agent waiting time is omitted; shorter runs hold their final frame. Native evaluator outcomes determine success.
The repository keeps the results, decision notes, protocol and four representative clips. The 600+ MB raw archive stays in the research workspace.
从真实相机画面和 action 记录出发，看看 3D waypoint 怎么来、动作怎么执行，以及换一台陌生机械臂时需要适配什么。
GITHUBStatic HTML, CSS and JavaScript plus the compact public evidence bundle.
RESULTSTask, seed, RoboDojo score, outcome, timing and recorded interaction counts.
METHODTask selection, allowed observations, motion tools, evaluation and prior-development disclosure.
Model setting. All 60 runs used GPT-6 Astra in xhigh mode, as confirmed by the experiment operator. The model setting itself was not serialized into every trial artifact, and the complete prompt/token ledger was not archived.
The button task had prior development. It had earlier source inspection and calibration. Removing those three runs changes the result from 11/60 to 8/57. The other runs also shared one adaptive conversation, so experience could carry across tasks.
This is a pilot, not a leaderboard submission. We used 20 selected tasks, three seeds per task and one public layout per seed. All runs are simulated; this experiment does not establish real-robot transfer.
About this project. The site comes from the completed codex-pilot20-v1 experiment record. Historical filenames still use Codex for the controller/bridge; the model making the decisions in this campaign was GPT-6 Astra in xhigh mode.
We used RoboDojo and its public implementation. This is an independent experiment and is not affiliated with the benchmark organizers.