logoHaga

Verification

Physics-violation detection: current benchmark results

·3 min read

100% detection of injected physics violations at zero false flags on clean runs — with the full methodology, thresholds, and confidence intervals published.

We are publishing our current benchmark results as raw numbers, not a case study. Every figure below is generated by the haga-core pipeline and traceable to versioned JSON artifacts (pillar1-*.json, pillar2-*.json, git SHA c848aab). If you want to verify any claim, ask for repo access.

What Haga checks

Two pillars:

  1. Policy stress gates — run a manipulation policy (Franka Panda) through nominal, mild, and moderate physics perturbation tiers; gate deployment on success-rate retention and grasp-failure deltas.
  2. Physics-consistency checking — score trajectories and AI-generated video against physical law: object permanence, ballistic motion, contact consistency, static hover, impulse limits.

Pillar 2: physics checker — synthetic violations

Injected violation classes, n=100 each:

Violation class Flagged Rate 95% CI
Teleportation 100/100 1.00 [0.963, 1.0]
Hover (unsupported objects) 99/99 1.00 [0.963, 1.0]
Impulse anomalies 100/100 1.00 [0.963, 1.0]
Penetration 100/100 1.00 [0.963, 1.0]

Negative controls (must NOT fire):

Cohort Flagged Rate
Clean trajectories 0/100 0.00
Clean + sensor noise 0/100 0.00

Recall 1.00, precision 1.00 on this suite. The negative controls matter more than the recall: a checker that cries wolf is worse than useless as a deployment gate.

Pillar 2: generative video — held-out cohort

Model: CogVideoX-5b-I2V, tracker CoTracker3, failure mode: static_hover.

Cohort n Flagged Rate 95% CI
Real reference footage 1 0 0.0 [0.0, 0.79]
CogVideoX generations 9 9 1.00 [0.66, 1.0]

All 9 held-out generations across 3 unseen seeds were flagged; the real clip passed clean. Small n — we publish it anyway, because a confidence interval that admits uncertainty is worth more than a round number that hides it.

Pillar 1: policy stress gates

Four manipulation tasks, 50 episodes per condition, Franka Panda:

Task Success (nominal) Success (mild perturbation) Gate verdict
Lift 1.00 1.00 PASS (score 1.00)
PickPlaceCan 1.00 1.00 PASS (score 1.00)
Stack 0.96 0.82 PASS (score 0.854)
Door 0.60 0.64 PASS (score 1.067)

The Door task shows what these gates are for: nominal success of 0.60 looks acceptable until severe-perturbation runs expose a 0.48 failure rate under load. A gate that only measures nominal success would have waved it through.

Honest scope statement

  • Results are simulation-based (Panda manipulation tasks, synthetic violation injection, one generative video model).
  • The checker's thresholds, check definitions, and scoring criteria are in the artifacts; we change them via versioned commits, never silently.
  • We have no customer pilot results to report yet. When pilots exist, they will be published here with the same artifact discipline — or not published at all.
  • Physical-robot validation is the next milestone, not a completed one.

If your team ships policies or world models into the real world, we'd like to run this suite against your stack.

Contact: haga@mushoodhanif.com.

Want an independent check on a policy or world-model artifact? Haga runs fixed protocols with shown failures — not self-graded demos.

Related reading