PhysEvo: Astra can act, let it
We introduce PhysEvo, an embodied self-evolution framework for the physical world, driven by execution feedback.
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.
- Eight tasks identified as VLA strengths in Table F2 of An Unexpected Robot Policy: Early Evaluations of GPT-6 Astra on RoboDojo and Beyond [2].
- Ten tasks from GPT 6 Astra as an Embodied Policy [1], selected by splitting the 0–72% range of official π0.5 success rates into four equal intervals and taking 6, 2, 1, and 1 tasks from lowest to highest.
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.
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.
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.
| Task | DM0.5[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 table | 44.00 | 46.33 | 57.67 | 62.50 | 23.33 | 47.50 | 45.00 | 60.00 | 30.00 | 39.50 | 60.00 |
| Classify by language | 0.47 | 1.07 | 2.00 | 1.33 | 0.60 | 3.87 | 3.27 | 38.00 | 60.00 | 46.00 | 100.00 |
| Imitate sorting sequence | 1.80 | 1.67 | 2.50 | 2.90 | 1.60 | 8.63 | 2.77 | 53.00 | 0.00 | 58.90 | 100.00 |
| Arrange the largest number | 7.85 | 4.11 | 8.56 | 4.36 | 2.29 | 15.75 | 11.48 | 50.00 | 57.00 | 68.70 | 100.00 |
| Pack objects into box | 14.72 | 17.12 | 18.69 | 20.83 | 18.36 | 22.40 | 23.93 | 50.00 | 50.00 | 11.60 | 70.00 |
| Classify objects | 26.83 | 10.33 | 17.20 | 5.53 | 24.67 | 19.73 | 20.67 | 71.00 | 100.00 | 69.00 | 100.00 |
| Build tower | 55.20 | 82.93 | 52.60 | 52.53 | 37.73 | 84.33 | 83.40 | 64.00 | 12.00 | 16.40 | 40.00 |
| Make Kong | 56.67 | 90.00 | 41.33 | 32.00 | 26.67 | 18.00 | 17.33 | 40.00 | 0.00 | 0.00 | 60.00 |
| Fold clothes | 28.96 | 32.75 | 41.49 | 51.31 | 29.12 | 37.65 | 34.83 | 100.00 | 40.00 | 76.40 | 68.00 |
| Put bottles in the dustbin | 81.70 | 96.30 | 97.70 | 94.03 | 79.93 | 97.90 | 98.00 | 100.00 | 36.00 | 35.30 | 100.00 |
| Insert tubes | 71.73 | 58.53 | 50.00 | 26.27 | 17.87 | 82.53 | 81.07 | — | — | 6.00 | 88.00 |
| Play tic-tac-toe | 35.77 | 65.23 | 36.80 | 64.07 | 8.23 | 95.67 | 94.53 | — | — | 10.70 | 100.00 |
| Pour balls into a vase | 9.33 | 28.00 | 46.00 | 17.33 | 13.33 | 30.67 | 26.67 | — | — | 4.00 | 80.00 |
| Pour liquid into a cup | 34.67 | 35.34 | 42.00 | 34.67 | 14.66 | 34.66 | 29.33 | — | — | 4.00 | 40.00 |
| Store laptop & headphones | 16.93 | 28.27 | 44.40 | 30.54 | 16.13 | 38.14 | 43.86 | — | — | 1.60 | 16.00 |
| Fill pen holder | 22.10 | 41.27 | 44.73 | 15.37 | 23.23 | 49.77 | 50.53 | — | — | 5.00 | 19.00 |
| Overall (10 tasks) | 31.82 | 38.26 | 33.97 | 32.73 | 24.43 | 35.58 | 34.07 | 62.60 | 38.50 | 42.18 | 79.80 |
| Overall (16 tasks) | 31.80 | 39.95 | 37.73 | 32.22 | 21.11 | 42.95 | 41.67 | — | — | 28.32 | 71.31 |
| Task | DM0.5[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 table | 4.00% | 11.33% | 14.67% | 13.33% | 0.00% | 11.00% | 7.00% | 0.00% | 0.00% | 0.00% | 0.00% |
| Classify by language | 0.00% | 0.00% | 0.67% | 0.00% | 0.00% | 1.00% | 1.00% | 20.00% | 40.00% | 30.00% | 100.00% |
| Imitate sorting sequence | 0.00% | 0.00% | 0.67% | 0.67% | 0.00% | 6.00% | 1.00% | 40.00% | 0.00% | 36.00% | 100.00% |
| Arrange the largest number | 3.20% | 0.53% | 3.73% | 0.53% | 0.53% | 8.20% | 5.40% | 40.00% | 40.00% | 60.00% | 100.00% |
| Pack objects into box | 1.33% | 2.93% | 2.40% | 3.73% | 2.40% | 7.20% | 9.40% | 20.00% | 20.00% | 0.00% | 40.00% |
| Classify objects | 16.67% | 4.00% | 8.67% | 1.33% | 12.67% | 9.00% | 11.00% | 60.00% | 100.00% | 56.00% | 100.00% |
| Build tower | 42.00% | 78.67% | 47.33% | 38.67% | 24.00% | 81.00% | 80.00% | 60.00% | 0.00% | 2.00% | 20.00% |
| Make Kong | 56.67% | 90.00% | 41.33% | 32.00% | 26.67% | 18.00% | 17.00% | 40.00% | 0.00% | 0.00% | 60.00% |
| Fold clothes | 24.53% | 28.00% | 38.93% | 49.07% | 21.07% | 36.40% | 33.00% | 100.00% | 40.00% | 72.00% | 60.00% |
| Put bottles in the dustbin | 72.00% | 94.00% | 96.67% | 91.33% | 69.33% | 97.00% | 97.00% | 100.00% | 20.00% | 10.00% | 100.00% |
| Insert tubes | 59.33% | 42.67% | 28.00% | 6.67% | 3.33% | 73.00% | 72.00% | — | — | 0.00% | 80.00% |
| Play tic-tac-toe | 3.33% | 40.00% | 5.33% | 58.67% | 1.33% | 89.00% | 86.00% | — | — | 0.00% | 100.00% |
| Pour balls into a vase | 9.33% | 28.00% | 46.00% | 17.33% | 13.33% | 31.00% | 27.00% | — | — | 4.00% | 80.00% |
| Pour liquid into a cup | 34.67% | 35.34% | 42.00% | 34.67% | 14.66% | 34.50% | 29.00% | — | — | 4.00% | 40.00% |
| Store laptop & headphones | 2.67% | 11.34% | 22.66% | 16.00% | 5.33% | 20.00% | 25.50% | — | — | 0.00% | 0.00% |
| Fill pen holder | 3.33% | 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% |
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.
Insert tubes
This recording could not be loaded. Use “Open clip” to check the video file.
Pour water
This recording could not be loaded. Use “Open clip” to check the video file.
06Baseline Comparison
PhysEvo against fourteen published baselines across five capability dimensions. Each cell is Score / success rate; Average is the equal-weight mean of the five dimensions.
| Rank | Model | Contributor | Average | Generalization | Precision | Long-horizon | Memory | Open |
|---|---|---|---|---|---|---|---|---|
| 🥇 | PhysEvo | Ours | 69.33/61.92% | 47.83/40.00% | 49.13/41.25% | 61.13/45.00% | 100.00/100.00% | 88.54/83.33% |
| 🥈 | Simate-beta | Simate | 33.95/27.96% | 35.09/27.95% | 34.35/26.92% | 57.84/43.42% | 33.33/33.00% | 9.12/8.50% |
| 🥉 | Liber-0 Preview | LiberAI | 30.74/25.52% | 24.99/18.61% | 38.28/33.17% | 45.98/33.17% | 37.77/37.33% | 6.68/5.33% |
| 4 | Liber-0 Lite | LiberAI | 29.24/24.23% | 25.36/19.11% | 36.60/31.08% | 45.50/32.92% | 35.68/35.44% | 3.06/2.58% |
| 5 | GPT-6-Astra | RoboDojo Team | 28.97/22.48% | 33.36/30.50% | 12.65/4.00% | 21.45/8.25% | 43.04/38.67% | 34.36/31.00% |
| 6 | DM0.5 | Dexmal | 24.90/19.34% | 15.77/10.95% | 24.82/16.75% | 33.70/19.50% | 47.74/47.44% | 2.43/2.08% |
| 7 | GalaxeaVLA (G0.5) | Galaxea AI | 20.23/14.88% | 18.46/12.83% | 28.25/20.42% | 44.12/32.25% | 8.61/7.33% | 1.73/1.58% |
| 8 | Xiaomi-Robotics-1 | Xiaomi Robotics | 20.07/13.93% | 23.54/17.00% | 26.69/18.83% | 38.39/23.67% | 7.81/6.56% | 3.94/3.58% |
| 9 | OpenWAM-α | OpenWAM Team | 17.18/11.92% | 20.71/14.83% | 18.45/9.25% | 34.93/25.33% | 10.41/9.11% | 1.41/1.08% |
| 10 | Meituan-Robotics-0 | Meituan Robotics | 14.95/9.53% | 13.75/8.17% | 16.77/7.75% | 29.61/18.58% | 10.06/8.89% | 4.54/4.25% |
| 11 | Hy-Embodied-0.5-VLA | Tencent Robotics X | 13.07/8.80% | 11.78/8.39% | 13.81/8.00% | 25.74/14.92% | 13.37/12.11% | 0.65/0.58% |
| 12 | KinRT | IIGroup | 13.02/8.80% | 14.02/8.61% | 15.65/9.92% | 26.40/18.08% | 4.82/3.56% | 4.23/3.83% |
| 13 | SimpleMemVLA | SimpleMemVLA Team | 12.58/9.27% | 6.36/3.95% | 7.42/2.92% | 14.58/5.50% | 33.71/33.22% | 0.85/0.75% |
| 14 | Spatial Forcing | OpenHelix Robotics | 12.38/8.04% | 14.12/9.34% | 17.32/10.58% | 23.26/14.58% | 5.43/4.11% | 1.78/1.58% |
| 15 | Pi-05 | RoboDojo Team | 11.44/6.93% | 13.38/8.17% | 12.40/5.50% | 23.54/14.67% | 5.89/4.67% | 1.98/1.67% |
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.