Get in touch

Physical AI

Validated and reliable robot policies.

Every motion is traced to what the policy sees.

Works on open-weight policies.

Built by researchers from Fraunhofer HHI.

Demo

From memorisation to generalisation.

In training, the plate always stood in the same spot, so the robot learned to focus on its location instead of looking at the plate. When we move the plate, it puts the bowl where the plate used to be. Layerwise shows it hardly looks at the plate. After retraining, it does.

InstructionPut the bowl on the plate.
OpenVLA-OFT

Memorisation failure

Relevance Maps by Layerwise
Loading video
Pick upPut on plate
OpenVLA-OFT fine-tuned by Layerwise

Generalisation

Relevance Maps by Layerwise
Loading video
Pick upPut on plate
Layerwise relevance map: the glow marks what the policy uses for its next movePlanned motionTrained spot of the plate

Offer

Shows what your policy relies on and which runs fail.

Included in every deployment

Relevance maps

Layerwise highlights the inputs that drive each motion.

  • For each camera and each phase of the task
  • Shown on your own logged rollouts

Failure detection

Layerwise reads the policy's internals and flags the runs that fail.

  • Shows which runs to review
  • Tested on tasks never seen in training

Optional service

Policy fine-tuning

We retrain your open-weight policy on the demonstrations it was missing.

  • Looks at the right objects after retraining

Proof

Results on three open robot policies.

Measured in simulation on OpenVLA-OFT, π0.5 and SmolVLA, against standard and published baselines.

Relevance maps

1st of 14

Most faithful relevance map among 14 attribution methods

OpenVLA-OFT on LIBERO

Faithfulness · three of the 13 others

Layerwise0.44
SmoothGrad0.41
Attention0.40
Integrated Gradients0.32

Failure detection

0.90

AUROC for predicting a failing run on tasks never seen in training

SmolVLA on LIBERO

Same test, other approaches

Layerwise0.90
Embedding distance0.84
Run length so far0.64
Outlier detection0.49

AUROC, 0.5 is chance. Best variant of each approach.

Reliable policy

+62%

More scenes with a moved plate solved after retraining on our demonstrations

OpenVLA-OFT on LIBERO · π0.5 gains as much

OpenVLA-OFT

Off the shelf0%
After retraining62%

π0.5

Off the shelf8%
After retraining69%

Pilot

Your policy, validated in eight weeks.

Together with our scientists, we explain your policy, set up monitoring and fix what it gets wrong.

  1. Week 1

    Connect

    We set up your checkpoint and replay your logged rollouts.

  2. Weeks 2 to 5

    Diagnose and report

    We show which inputs drive your policy, and which runs fail. You decide on rollout.

  3. Weeks 6 to 8

    Fine-tune, optional

    A retrained policy with measured results.

  4. After week 8

    Ongoing support

    In production, and when your tasks or your robot change.

What we need from you

  • One checkpoint of an open-weight policy

    Such as OpenVLA-OFT or π0.5.

  • Logged rollouts of one task

    From your own robot.

  • One engineer as a contact

    For access and questions on the setup.

What you get

  • What the policy looks at

    On your own rollouts.

  • A list of memorised motions

  • A retrained policy, optional

    With the measured result.

Let's validate your policy now.

Tell us which policy you run and which tasks your robot should do.

Get in touch