Does sim red-teaming predict real-robot behaviour?
The first objection to any simulation-only safety tool is "but it's only sim." This page is the honest answer, with citations — not a promise.
The short version
Provael is a pre-deployment scanner: it measures how often an attack drives a policy out of its safe envelope in simulation, before that policy ever touches hardware. The relevant research finding is that, for learned visuomotor policies, sim (and even edited-image) evaluation predicts real-robot failure well enough to be useful — without breaking real robots to find out. The strongest evidence below is about diffusion policies rather than VLAs; the VLA evidence is real but weaker, and that difference is marked per citation rather than averaged away.
- Predictive Red Teaming (Majumdar et al., 2025) degrades a policy's inputs (lighting, textures, distractors, camera pose) in sim / on edited images and shows the predicted per-factor success tracks real-robot success closely — in the paper's own words, "less than 0.19 average difference between predicted and real success rates", across 500+ hardware trials in twelve off-nominal conditions. It identifies brittle factors without real-world rollouts. This result is on visuomotor diffusion policies, not VLAs. It is the strongest sim-predicts-real evidence on this page and it is not about the model class Provael tests, so it is a reason to think the approach is sound rather than a measurement about VLAs. arXiv:2502.06575 · https://arxiv.org/abs/2502.06575
- Robustness benchmarks (LIBERO-Plus, LIBERO-PRO, 2025–26) independently find VLAs collapse under camera/position shifts and largely ignore language perturbations — the same failure modes Provael's instruction/visual families exercise. This one is about VLAs, but it measures robustness rather than sim-to-real agreement — it corroborates the failure modes, not the prediction claim. arXiv:2510.13626
- Real-to-sim evaluation (SimplerEnv) is the methodology VLA papers use precisely because sim rankings track real hardware closely enough to compare policies. https://simpler-env.github.io/
What this does and does not claim
- Does: give a reproducible, controlled, pre-deployment signal (calibrated ASR + 95% CI + benign-FPR control) that flags brittle policies early and cheaply, and that the literature says correlates with real-robot brittleness.
- Does not: certify real-world safety, predict a specific robot's behaviour on a specific day, or replace hardware testing and functional-safety processes. Treat the ASR as a floor on susceptibility, not a guarantee. See compliance (evidence, not certification) and the README's "Scope and honest limitations."
How to keep your own results honest
- Always run the
nonebenign baseline and read every attack rate against that control. - Calibrate the predicate to a benign-FPR target (
provael calibrate) so "unsafe" means "diverted past the policy's own benign envelope," not an arbitrary threshold. - Report CIs and
n, not just point estimates —n = 10per attack is a screen, not a proof.
The rule for any cross-simulator or sim-to-real claim
No cross-simulator or sim-to-real claim without per-robot trajectory calibration, and without the within-simulator seed spread as the yardstick. A rate measured in one simulator and a rate measured in another, or on a robot, differ for reasons that have nothing to do with the policy unless the two are tied together, robot by robot, by calibrating the trajectories they produce. And a single run is one draw: two runs of one configuration in one simulator already differ, so a cross-simulator or sim-to-real difference that sits inside that spread says nothing.
The results READMEs print the yardstick. Each multi-task run carries a per-seed table, each arm's
unsafe count pooled over tasks with its range across seeds, and where two committed runs share a
configuration, the run-to-run difference with its interval. On SmolVLA x LIBERO-Object the two
runs of the published configuration, on provael 0.32.0 and 0.41.2, put roleplay at 44/50 and
42/50, a difference of -4 points with a 95% interval of [-18, +10]
(the 14 September run).
No committed result runs a real policy in a second simulator or on a robot (results/hardware/
holds no run), so Provael makes no such claim.