PhysEvo Research notes
Robot manipulation · Research preview · Sep 2026

PhysEvo: Astra can act, let it

We introduce PhysEvo, an embodied self-evolution framework for the physical world, driven by execution feedback.

19 September 2026Contact: johniq213@gmail.com
Insert tubes — task illustration 9752
Official Astra and our task scores 050100 RoboDojo official GPT-6-Astra · ScoreAstra+ PhysEvo6.0088.00
Insert tubes
Data: Astra: RoboDojo [2] · + PhysEvo: Ours.

PhysEvo scores 71.31 across all sixteen RoboDojo tasks. Explore the full results, followed by the ten-task and eight-task comparisons.

01Overall Performance on 16 Tasks

These sixteen tasks come from two sources.

PhysEvo achieves an overall score of 71.31. Baseline results across all sixteen tasks come from the RoboDojo Leaderboard [3].

[1] does not report the eight Table F2 tasks, so its results are not included in the sixteen-task overall.

FIG. 01Comparison across all sixteen RoboDojo tasks.Data: RoboDojo leaderboard [3] — all comparison methods, sixteen tasks (17 Sep 2026). Ours — PhysEvo.

0210 Tasks from [1]

On the ten tasks from GPT 6 Astra as an Embodied Policy [1], PhysEvo scores 79.80, alongside 62.60 for Hybrid and 38.50 for Direct.

FIG. 02Ten-task comparison.Data: Hybrid report [1] — Astra + π0.5 Hybrid, Astra Direct and reproduced VLA baselines. RoboDojo leaderboard [3] — Liber-0 Preview / Lite (17 Sep 2026). Ours — PhysEvo.

038 Tasks from Table F2 in [2]

These eight tasks were identified as VLA strengths in Table F2 of An Unexpected Robot Policy: Early Evaluations of GPT-6 Astra on RoboDojo and Beyond [2]. We retain that task set and compare each method’s mean score.

FIG. 03Comparison on the eight Table F2 tasks. Both views retain Score ordering.Data: RoboDojo leaderboard [3] — all comparison methods (17 Sep 2026). Ours — PhysEvo. Task set: Official Astra report [2], Table F2.

04Per-Task Performance

PhysEvo also improves fine manipulation performance compared with Astra Direct. On Build tower, the score increases from 12.00 to 40.00; on Make Kong, it increases from 0.00 to 60.00.

Score 0100
TaskDM0.5[1][3]G0.5[1][3]Xiaomi
Robotics-1
[1][3]
OpenWAM-α[1][3]π0.5[1][3]Liber-0
Preview
[3]
Liber-0
Lite
[3]
Astra + π0.5
Hybrid
[1]
Astra Direct[1]Official
Astra
[2]
Ours
PhysEvo
Organize table44.0046.3357.6762.5023.3347.5045.0060.0030.0039.5060.00
Classify by language0.471.072.001.330.603.873.2738.0060.0046.00100.00
Imitate sorting sequence1.801.672.502.901.608.632.7753.000.0058.90100.00
Arrange the largest number7.854.118.564.362.2915.7511.4850.0057.0068.70100.00
Pack objects into box14.7217.1218.6920.8318.3622.4023.9350.0050.0011.6070.00
Classify objects26.8310.3317.205.5324.6719.7320.6771.00100.0069.00100.00
Build tower55.2082.9352.6052.5337.7384.3383.4064.0012.0016.4040.00
Make Kong56.6790.0041.3332.0026.6718.0017.3340.000.000.0060.00
Fold clothes28.9632.7541.4951.3129.1237.6534.83100.0040.0076.4068.00
Put bottles in the dustbin81.7096.3097.7094.0379.9397.9098.00100.0036.0035.30100.00
Insert tubes71.7358.5350.0026.2717.8782.5381.07——6.0088.00
Play tic-tac-toe35.7765.2336.8064.078.2395.6794.53——10.70100.00
Pour balls into a vase9.3328.0046.0017.3313.3330.6726.67——4.0080.00
Pour liquid into a cup34.6735.3442.0034.6714.6634.6629.33——4.0040.00
Store laptop & headphones16.9328.2744.4030.5416.1338.1443.86——1.6016.00
Fill pen holder22.1041.2744.7315.3723.2349.7750.53——5.0019.00
Overall (10 tasks)31.8238.2633.9732.7324.4335.5834.0762.6038.5042.1879.80
Overall (16 tasks)31.8039.9537.7332.2221.1142.9541.67——28.3271.31
FIG. 04Scores across all sixteen tasks. — indicates an unreported result.Data: GPT 6 Astra as an Embodied Policy [1] — Hybrid, Direct and VLA results for the first ten tasks. RoboDojo leaderboard [3] — VLA results for the remaining six tasks and all Liber-0 results (17 Sep 2026). An Unexpected Robot Policy: Early Evaluations of GPT-6 Astra on RoboDojo and Beyond [2] — Official Astra. Ours — PhysEvo.
Success rate / coverage 0100
TaskDM0.5[1][3]G0.5[1][3]Xiaomi
Robotics-1
[1][3]
OpenWAM-α[1][3]π0.5[1][3]Liber-0
Preview
[3]
Liber-0
Lite
[3]
Astra + π0.5
Hybrid
[1]
Astra Direct[1]Official
Astra
[3]
Ours
PhysEvo
Organize table4.00%11.33%14.67%13.33%0.00%11.00%7.00%0.00%0.00%0.00%0.00%
Classify by language0.00%0.00%0.67%0.00%0.00%1.00%1.00%20.00%40.00%30.00%100.00%
Imitate sorting sequence0.00%0.00%0.67%0.67%0.00%6.00%1.00%40.00%0.00%36.00%100.00%
Arrange the largest number3.20%0.53%3.73%0.53%0.53%8.20%5.40%40.00%40.00%60.00%100.00%
Pack objects into box1.33%2.93%2.40%3.73%2.40%7.20%9.40%20.00%20.00%0.00%40.00%
Classify objects16.67%4.00%8.67%1.33%12.67%9.00%11.00%60.00%100.00%56.00%100.00%
Build tower42.00%78.67%47.33%38.67%24.00%81.00%80.00%60.00%0.00%2.00%20.00%
Make Kong56.67%90.00%41.33%32.00%26.67%18.00%17.00%40.00%0.00%0.00%60.00%
Fold clothes24.53%28.00%38.93%49.07%21.07%36.40%33.00%100.00%40.00%72.00%60.00%
Put bottles in the dustbin72.00%94.00%96.67%91.33%69.33%97.00%97.00%100.00%20.00%10.00%100.00%
Insert tubes59.33%42.67%28.00%6.67%3.33%73.00%72.00%——0.00%80.00%
Play tic-tac-toe3.33%40.00%5.33%58.67%1.33%89.00%86.00%——0.00%100.00%
Pour balls into a vase9.33%28.00%46.00%17.33%13.33%31.00%27.00%——4.00%80.00%
Pour liquid into a cup34.67%35.34%42.00%34.67%14.66%34.50%29.00%——4.00%40.00%
Store laptop & headphones2.67%11.34%22.66%16.00%5.33%20.00%25.50%——0.00%0.00%
Fill pen holder3.33%18.67%22.00%6.00%7.33%25.00%31.00%——0.00%0.00%
Overall (10 tasks)22.04%30.95%25.51%23.07%15.67%27.48%26.18%48.00%26.00%26.60%68.00%
Overall (16 tasks)20.82%30.34%26.32%23.13%12.62%34.21%33.27%——17.12%61.25%
FIG. 05Completion across all sixteen tasks. — indicates an unreported result.Data: GPT 6 Astra as an Embodied Policy [1] — Hybrid, Direct and VLA results for the first ten tasks. RoboDojo leaderboard [3] — VLA results for the remaining six tasks and all Liber-0 results (17 Sep 2026). RoboDojo leaderboard [3] — Official Astra. Ours — PhysEvo.

05Robots in motion.

Recordings from two sources: simulation rollouts across sixteen RoboDojo tasks, each with synchronized overview and wrist-camera views, and demonstrations on a physical robot.

SIMULATION · 16 TASKS

Insert tubes

0:14 Open clip

REAL ROBOT · 5 DEMONSTRATIONS

Pour water

0:12 Open clip

07Conclusion

PhysEvo shows that Astra can act — including on the fine manipulation that prior work treats as the home turf of specialized VLAs. What enables this is a pairing: the physical environment in which the model acts, and the self-evolution framework that turns execution feedback into cumulative improvement — both of which are essential. This more human-like generalization will undoubtedly deal a fundamental blow to the embodied technology routes that came before it.