Non-Profit Robot Safety & Robustness Research
Physical safety & robustness for learned robot control.
An independent research laboratory dedicated to safety, out-of-distribution robustness, and verifiable physical grounding across learned robot policies — from Vision-Language-Action (VLA) foundation models to diffusion policies and end-to-end neural controllers. We develop zero-parameter runtime safety layers and diagnostic standards that prevent catastrophic policy collapse and unsafe physical execution under real-world perturbations.
End-to-end learned policies — whether large VLA foundation models, diffusion policies, or imitation learning networks — are brittle: small shifts in camera angle, lighting, or initial robot pose cause catastrophic policy collapse and unsafe trajectories. Distortion Lab builds external safety layers, diagnostic audits, and robust control architectures that guarantee physical reliability by construction.
Protecting trained robot policies — including VLAs (OpenVLA, π₀, Octo), Diffusion Policies, ACT, and neural controllers — from real-world sensory noise, visual shifts, and initial configuration drift without expensive retraining.
Control-theoretic external wrappers and joint-space Jacobian steering layers that enforce physical limits and prevent compounding drift before policy actions can drive manipulators into unsafe configurations or damage hardware.
Systematic safety probing across benchmark suites (LIBERO-Plus, RoboCasa, SimplerEnv). Decomposing policy failures into perceptual vs. kinematic causes to diagnose why end-to-end trained models break outside their training distribution.
Verifying safety layers and lightweight residual compensators on physical 7-DoF manipulators (Franka Emika Panda) and low-cost hardware, ensuring that simulation robustness guarantees hold during physical execution.
Compare twenty-four steps of open-loop execution from a perturbed initial pose against analytic Jacobian return (return-then-release). Trajectories run directly in your browser using the lab's canvas kinematics engine.
The baseline drifts because actions were predicted for a pose the arm is not in. Pausing the policy for 24 steps while a Jacobian controller returns the arm to its canonical configuration restores success to 97.5% (and 96.7% with learned imitation).
The hardest perturbation column of LIBERO-Plus (robot initial state) is commonly diagnosed as a perception breakdown. We show that the collapse is purely kinematic: the arm starts outside the narrow support of demonstrations. Pausing the policy for 24 steps while a training-free Jacobian controller returns the arm to its canonical configuration lifts OpenVLA-OFT+ from 30.4% to 97.5% [94.7, 98.8] (p = 0.004, 166 fixed / 5 broken) at zero cost on clean vanilla LIBERO (+0.5 pp). A 135k-parameter learned imitation matches it at 96.7% (R² = 0.996), and a gripper-pose return reaches 95.0% without joint encoders.
@article{kravchenko2026robot,
title={Robot Initial State Is Not a Robustness Problem: Return, Then Release, Around a Frozen VLA Policy},
author={Kravchenko, Philip},
journal={arXiv preprint},
year={2026},
institution={Distortion Lab},
url={https://distortionlab.org/paper.html}
}All code, model weights, and per-episode evaluation logs are made public under permissive open-access licenses. Anyone can reproduce, verify, and extend our results without commercial barriers.
Complete reproducible test runner for LIBERO-Plus, Jacobian projection module, deterministic seed harness, and automated regression test suite.
View RepositoryLightweight 590k parameter residual head weights for OpenVLA-OFT and VLA-Adapter. Ready for zero-shot deployment in standard robot pipelines.
Download CheckpointsComplete per-episode JSONL logs covering all 1,550 evaluated variants across three policies. Recompute all figures without executing models.
Inspect Raw TelemetryWhich evaluation protocol, control mode, seed count, and exact git commit. A metric without its protocol is not a result.
Success rates ship with Wilson 95% confidence intervals. An improvement smaller than two standard deviations is noise, not progress.
Perturbed and unperturbed clean performance in the same table. Robustness bought by degrading nominal capability is documented explicitly.
Fixed in writing with a timestamp prior to execution. Declared after the run, they are worthless as empirical evidence.
Task by task, success or failure, for every seed, including all failed episodes. External researchers can recompute our tables without running GPU hours.
Report absolute scores against leaderboards and effects against internal baselines. Never substitute cross-harness subtractions, which drift by points.
Negative results are reported with failure mechanisms. The negative results are what make a positive result meaningful.
The work is the grounding layer. An intervention requiring retraining foundation checkpoints breaks the thesis and must be flagged as an exception.
Every time something in this lab reported success while doing nothing or the wrong thing, it was documented alongside the check that now prevents it. Most failures occurred in the evaluation harness and cloud infrastructure. That machinery now answers to the same standard as peer-reviewed analysis.
What reported success: A queue passed a flag to a collector that predated it. Unknown settings are ignored by design, so two machines re-collected the wrong data for three hours.
The check it produced: A run first asserts that the code on the far end has every flag it is being sent, and refuses before spending anything.
What reported success: A code patch matched its anchor as a substring and landed at the wrong indentation — twice, the second time after the lesson had been written down.
The check it produced: Anchors include the leading newline, so they match a whole line or nothing.
What reported success: A file filter meant to skip 16.5 GB of merged weights also skipped the trained adapter itself. A more forgiving library would have shipped the base model under our name.
The check it produced: The merge refuses an adapter directory with no weights in it, and the bundle’s size is asserted before it moves.
What reported success: A rented machine passed the usual GPU check and could not allocate a single tensor. The worker ran for four minutes and wrote its completion marker anyway.
The check it produced: Provisioning ends with a real allocation and real arithmetic on the device, and no script writes a marker without checking its own exit status.
What reported success: The evaluator printed “reporting split” in its header and filtered nothing: a trained module would have been scored on the tasks it was fitted to.
The check it produced: The split is enforced as an exclusion, and the run refuses when it excluded everything and when it excluded nothing.
What reported success: A caption said three arms were measured on the same tasks. Two of them had been run on a different policy and a different set of variants.
The check it produced: “Same tasks” is a set equality: a test asserts identical keys before it checks a single number.
What reported success: Pooling four task suites inverted a headline statistic: a correlation read 0.085 pooled and +0.55 to +0.69 inside every suite that could be pooled at all.
The check it produced: Before pooling anything, check that the quantity means the same thing in each group. Here it is one line, living in the computing code.
What reported success: A pre-flight check printed “OK” and then hung for twenty-one minutes inside the graphics driver, on a machine that could not render a frame.
The check it produced: The render runs in a subprocess under a ninety-second cap. A failure is a decision; a hang is not.
What reported success: A run finished cleanly. Hours later the rented machine exited on its own and could not be restarted, with ninety minutes of finished measurement on its disk.
The check it produced: Each shard is pulled home the moment it finishes. A result that exists only on a rented disk is not a result yet.
The cure has never been more care. It is a check that fails loudly — and the cheap, decisive check goes before the expensive work, not after it.
Distortion Lab is an independent, non-profit research laboratory dedicated to robustness, safety, and physical grounding across learned robot control systems. We investigate why trained neural policies — from VLA foundation models to diffusion and imitation policies — fail under real-world perturbations, and build verifiable runtime safety layers that make autonomous physical systems reliable for deployment. All our work is published openly without paywalls, commercial exclusivity, or closed APIs.
If you work on embodied policies, foundation model evaluation, or want the raw logs behind any published figure — or if you believe any of our numbers are in error — challenging our measurements is the most valuable contribution to this work.
Distortion Lab · Independent research in embodied AI, kinematic grounding, and evaluation. Direct inquiries: lab.distortion@gmail.com. Every metric on this page stems directly from dated commits and per-episode execution logs.
© 2026 Distortion Lab
Set in Big Shoulders Display, Archivo, and Martian Mono. Compliant with WCAG AA accessibility standards.