# RoboPro-Baselines

Evaluation of three base (non-DA3) VLA policies — **PI0.5**, **PI0** and **X-VLA** — on the RoboPRO
benchmark. All metrics are percentages, micro-averaged over episodes.

**SR** = success rate  ·  **HSR** = hard success rate (success **and** no collision)  ·  **CR** = collision rate

## What's in here

| Study | Scope | Where |
|---|---|---|
| **Full perturbation sweep** | 13 axes x 12 tasks x clean+clutter, all 3 models (5,148 cells) | [`perturbation_full/`](perturbation_full/) |
| **Baseline task suite** | 79 tasks x 4 scenes, clean and clutter, all 3 models | [`<model>/clean/`](pi05/clean/), [`<model>/clutter/`](pi05/clutter/) |
| **PI0.5 data composition** | 100 demos/task split 3 ways + default baseline | [`pi05/data_variants/`](pi05/data_variants/) |
| **Half-subset baselines** | 50 clean + 50 clutter per task, 12 tasks | [`half_models/`](half_models/) |
| **Collision breakdown** | floor vs object collision rates | [`collision_breakdown/`](collision_breakdown/) |
| **Checkpoints** | trained model weights | `pi05-roboreal-half-20k/`, `pi0-roboreal-half-40k/`, `xvla-roboreal-half-60k/` |

## Headline: robustness under perturbation (12 tasks)

Non-perturbed baseline vs the full 13-axis perturbation sweep, on the same 12 tasks.

| Model | Baseline clean | Perturbed clean | Baseline clutter | Perturbed clutter |
|---|---:|---:|---:|---:|
| **PI0.5** | 78.8 | 74.7 | 51.3 | 52.3 |
| **PI0** | 65.4 | 56.5 | 34.6 | 41.0 |
| **X-VLA** | 51.7 | 44.9 | 30.4 | 32.5 |

SR shown. **PI0.5 > PI0 > X-VLA holds on every axis and in both splits.** Perturbation costs each model
4-9 SR points on clean, as expected.

The clutter columns behave differently and should be read with care: perturbed clutter scores *higher*
than baseline clutter for all three models (PI0.5 51.3 -> 52.3, PI0 34.6 -> 41.0, X-VLA 30.4 -> 32.5).
Perturbation does not make tasks easier. The two campaigns build cluttered scenes from different
configs — the baseline uses `bench_demo_<scene>_<density>` while the perturbation sweep uses
`bench_perturb_*` / `bench_verified_*` — so their clutter is not the same distribution and the clutter
columns are **not** directly comparable across campaigns. Within a single campaign the comparisons are
sound; the clean columns, which share a configuration, are the ones to compare across campaigns.

## Perturbation by axis group (SR, clean / clutter)

> **Note on X-VLA Object-OOD:** the campaign was concluded with 79 X-VLA object-OOD cells
> unevaluated, so that one cell of the table below is computed over a task subset and is not
> matched to the other two models. Every other figure on this page is complete.


| Axis group | PI0.5 | PI0 | X-VLA |
|---|---:|---:|---:|
| Language (3 axes) | 68.3 / 46.5 | 51.2 / 36.9 | 38.8 / 28.6 |
| Distractor (2) | 70.2 / 46.5 | 53.5 / 33.8 | 35.6 / 28.8 |
| Vision (4) | 78.6 / 58.5 | 57.3 / 47.0 | 52.1 / 35.2 |
| Object-OOD (4) | 78.3 / 53.4 | 61.4 / 41.7 | 47.3 / 35.4 |

Language and distractor perturbations hurt most, vision least — the same ordering for all three models,
pointing at instruction grounding rather than visual robustness as the weak point.

## Baseline task suite (79 tasks, per model)

| Model | Clean SR / HSR / CR | Clutter SR / HSR / CR (avg d6-d15) |
|---|---|---|
| **PI0.5** | 70.2 / 65.7 / 11.9 | 60.9 / 47.8 / 31.7 |
| **PI0** | 60.1 / 55.9 / 14.7 | 47.3 / 32.9 / 44.6 |
| **X-VLA** | 49.5 / 46.7 / 9.4 | 39.6 / 27.5 / 38.7 |

Per-scene and per-task breakdowns are in `<model>/clean/<model>_clean.csv` and
`<model>/clutter/<model>_clutter.csv`.

## PI0.5 training-data composition (12 tasks)

| Training data | Demos/task | Clean SR | Clutter SR |
|---|---|---:|---:|
| default — full dataset | full | **78.8** | 51.3 |
| 50 clean + 50 clutter | 100 | 74.2 | **62.9** |
| 100 clean only | 100 | 71.2 | 23.8 |
| 100 clutter only | 100 | 66.7 | 55.4 |

A 50/50 mix beats the full-dataset default on clutter by +11.6 SR using only 100 demos per task.
Clean-only training collapses in clutter (23.8 SR, 79.2 CR). Full discussion in
[`pi05/data_variants/`](pi05/data_variants/).

## Directory layout

```
perturbation_full/                     # full 13-axis sweep: metrics, cell audit, methodology
<model>/clean/<model>_clean.csv        # 79-task baseline, obstacle-free
<model>/clutter/<model>_clutter.csv    # 79-task baseline, densities d6-d15 averaged
<model>/perturbation/                  # per-model copy of the 13-axis sweep
    <model>_perturbation.csv           #   per axis/scene/task/condition
    <model>_perturbation_averages.csv  #   scene-level aggregates
<model>/<group>/<scene>/<task>/
    _episodes.jsonl                    # per-seed rows, each tagged with its config
    <config>_seed<seed>.mp4            # rollout video
pi05/data_variants/                    # training-data composition study
```

## Reading the numbers

**Use HSR, not SR, when comparing under clutter.** All three models lose ~12 SR points when collisions
are counted as failures (X-VLA 12.1, PI0 12.6, PI0.5 12.0), with collision rates of 48-53%. Roughly a
quarter of every model's cluttered 'successes' involve pushing through an obstacle, and no model is
meaningfully better at avoidance than the others — PI0.5 leads by succeeding more often, not more gently.

A small number of cells are incomplete and are excluded from all averages above; 27 of 5,148 cells
(0.52%) are permanently short by one rollout for documented hardware reasons. See
[`perturbation_full/README.md`](perturbation_full/README.md) and `accepted_gaps.json` for the per-cell record.