RageVLA: Augmentation-Trained Perception and Action for State-of-the-Art Robustness on LIBERO
Abstract
Vision-language-action (VLA) models look saturated on LIBERO, yet most of them fall apart on LIBERO-PRO as soon as an object moves or the instruction asks for something new. We argue that what is missing is not scale but decoupling. In these policies, what to do is tied to where things were during training, to the scene around them and to the wording of the instruction. We cut these ties in two places. In the data, LIBERO demonstrations are re-arranged by general rules: every object and fixture is moved and rotated, tasks are carried into other environments and onto other surfaces, and every instruction is paraphrased. In the policy, a perception engine on frozen DINOv3 features, trained on this augmented data, hands the action model the scene in decoupled form: a list of objects with their positions, and a compact code of what the gripper is looking at, next to image and instruction embeddings from TinyCLIP. A flow-matching diffusion transformer, trained from scratch, acts on this view. We call this combination of Robust Augmentation and Grounded Encoding RageVLA. With 0.21 billion parameters, RageVLA sets a new state of the art on LIBERO-PRO among VLA policies, scoring 0.72 against 0.53 for the best published model, and a larger variant reaches 0.73. Two further tests on LIBERO-Object show that the policy acts on what it perceives. When the object positions it is given are shuffled while the camera images stay correct, its success falls from 0.965 to 0.215. When the target object is displaced by 28 cm, it still succeeds in 0.865 of the episodes, where , run through the same harness, succeeds in 0.105.
Figures & tables
| Split | Trajectories | Samples |
|---|---|---|
| Training | 132,610 | 2.79M |
| Held-out test | 1,751 | 36.6k |
| Validation | Held-out test | |
|---|---|---|
| Objects found | 99.3% | 99.5% |
| Object position error, median | 0.26 cm | 0.27 cm |
| Wrist cell class accuracy | 99.8% | |
| Wrist object cell accuracy | 99.4% | |
| Wrist depth error, median | 0.31 cm | |
| LIBERO-Goal | LIBERO-Spatial | LIBERO-10 | LIBERO-Object | ||||||||||||||||||
| Model | Obj | Pos | Sem | Task | Env | Obj | Pos | Sem | Task | Env | Obj | Pos | Sem | Task | Env | Obj | Pos | Sem | Task | Env | Total |
| OpenVLA | 0.96 | 0.00 | 0.98 | 0.00 | 0.98 | 0.97 | 0.00 | 0.97 | 0.00 | 0.89 | 0.81 | 0.00 | 0.96 | 0.00 | 0.85 | 0.98 | 0.00 | 0.98 | 0.00 | 0.00 | 0.52 |
| 0.94 | 0.00 | 0.93 | 0.00 | 0.39 | 0.95 | 0.00 | 0.97 | 0.00 | 0.60 | 0.79 | 0.00 | 0.82 | 0.00 | 0.27 | 0.94 | 0.00 | 0.90 | 0.00 | 0.29 | 0.44 | |
| 0.97 | 0.38 | 0.97 | 0.00 | 0.46 | 0.97 | 0.20 | 0.97 | 0.01 | 0.46 | 0.92 | 0.08 | 0.93 | 0.01 | 0.46 | 0.98 | 0.17 | 0.96 | 0.01 | 0.73 | 0.53 | |
| MolmoAct | 0.68 | 0.00 | 0.85 | 0.00 | – | 0.90 | 0.00 | 0.88 | 0.00 | – | 0.54 | 0.00 | 0.74 | 0.06 | – | 0.92 | 0.06 | 0.96 | 0.00 | – | 0.41 |
| NORA | 0.58 | 0.00 | 0.88 | 0.00 | – | 0.92 | 0.00 | 0.91 | 0.00 | – | 0.46 | 0.00 | 0.74 | 0.00 | – | 0.86 | 0.00 | 0.92 | 0.00 | – | 0.40 |
| Model | Params | Total | Obj+Sem | Pos+Task | Pos+Task+Env |
|---|---|---|---|---|---|
| OpenVLA | 7.50B | 0.52 | 0.95 | 0.00 | 0.23 |
| 3.30B | 0.44 | 0.90 | 0.00 | 0.13 | |
| 3.30B | 0.53 | 0.96 | 0.11 | 0.25 | |
| MolmoAct | 8.10B | 0.41 | 0.81 | 0.01 | – |
| NORA | 3.80B | 0.40 | 0.78 | 0.00 | – |
| X-VLA | 0.90B | 0.46 | 0.88 | 0.04 | – |
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
| RageVLA s1 | RageVLA n1 | |
| Width, heads | 384, 6 | 512, 8 |
| Denoiser blocks | 12 | 20 |
| Parameters | 35M | 100M |
| Peak learning rate | ||
| Warm-up steps | 500 | 2,000 |
| Epochs | 23 | 18 |
| Goal | Spatial | LIBERO-10 | Object | Mean | |
|---|---|---|---|---|---|
| s1 | 0.96 | 0.96 | 0.95 | 0.98 | 0.96 |
| n1 | 0.97 | 0.95 | 0.92 | 0.98 | 0.96 |
| RageVLA s1 | ||||||
| Axis | Shift | All | Feasible | All | Feasible | Infeasible |
| 7 cm | 0.980 | 0.980 | 0.910 | 0.910 | 0 | |
| 14 cm | 0.880 | 0.880 | 0.595 | 0.595 | 0 | |
| 21 cm | 0.905 | 0.905 | 0.395 | 0.395 | 0 | |
| 28 cm | 0.865 | 0.865 | 0.105 | 0.105 | 0 | |
| 35 cm | 0.445 | 0.830 | 0.000 | 0.000 | 100 | |
| Swap (leaderboard Pos) | Graded shift | |
|---|---|---|
| Suites | all four (40 tasks) | LIBERO-Object (10 tasks) |
| What moves | 2–4 objects exchange regions | picked object only |
| Magnitude | fixed by the object pair | 7, 14, 21, 28, 35 cm; or |
| Reported in | leaderboard, Table III | Fig. 4 of [ 5 ] |
| Used here | Table III | Sec. V-D |
| Goal | Spatial | LIBERO-10 | Object | Mean | |
|---|---|---|---|---|---|
| , leaderboard | 0.00 | 0.01 | 0.01 | 0.01 | 0.01 |
| , BDDL instruction [ 29 ] | 0.22 | 0.53 | 0.18 | 0.10 | 0.26 |
| RageVLA s1, BDDL instruction | 0.20 | 0.51 | 0.42 | 0.96 | 0.52 |
Robustness of VLAs
Robustness of vision-language-action (VLA) models concerns reliable robot behavior under changes to language, vision, sensors, tasks, and hardware.
See all 239 papers