← PORTFOLIO RÉSUMÉ
CASE STUDY / 04

After the Shield

Research into how a robot’s next proposed action changes after a safety system corrects its previous action. The project is in simulator and model setup; the paired experiments have not produced results yet.

TYPEVLA safety research
STATUSSprint 0 / bring-up
BASEπ₀.₅-LIBERO
BENCHMARKSafeLIBERO

01 / QUESTION

Does a safety intervention change what the policy wants to do next?

Runtime shields can prevent an immediate unsafe action, but that correction also changes the physical state the robot reaches. A learned VLA then observes a different world than it would have seen without the intervention. The project asks whether that changed state causally alters the policy's next nominal action proposal.

PRIMARY QUESTION

Downstream policy effect

Measure whether one safety correction changes the VLA's next nominal proposal.

CONTROL

Same state, paired branches

Both branches start from the exact same pre-intervention simulator state and synchronized policy randomness.

SEPARATION

Fresh-policy vs stale chunk

Distinguish a changed-state causal effect from deployment issues caused by cached action chunks.

COMPARISON

Defense architecture

Compare reactive AEGIS with an anticipatory/hybrid Critic-VLA arm after the causal pilot works.

02 / CAUSAL DESIGN

One intervention state. Two counterfactual futures.

At a genuine safety-intervention timestep s_t, the simulator state is snapshotted and policy randomness is synchronized. The experiment then branches from that identical point.

BRANCH N / NO INTERVENTION

Execute the nominal action.

  1. Restore identical s_t.
  2. Restore/synchronize policy RNG.
  3. Execute a_nom(t).
  4. Observe s_N(t+1).
  5. Force a fresh policy query.
record a_nom^N(t+1)
SAME START ↙ ↘
BRANCH S / SINGLE INTERVENTION

Execute one shield-corrected action.

  1. Restore identical s_t.
  2. Restore/synchronize policy RNG.
  3. Execute a_safe(t) once.
  4. Observe s_S(t+1).
  5. Force a fresh policy query with correction disabled for measurement.
record a_nom^S(t+1)
PRIMARY ENDPOINT D1 = L2(a_nom^S(t+1) − a_nom^N(t+1)) A paired next-proposal divergence, not an unpaired episode average.

03 / SCIENTIFIC CONTROLS

The experiment semantics are frozen before results.

The research repo treats experiment definitions as controlled artifacts. Primary semantics, exclusions, metrics, pairing rules, and go/no-go conditions are documented before scaled experiments.

Freeze the primary endpoint The next-policy divergence metric is defined before the causal pilot and cannot be silently changed after observing outcomes.
Log every intervention-relevant state Per-timestep schema includes nominal action, executed action, intervention magnitude, simulator-state reference, seed, collision, safety violation, and task outcome.
Predefine valid exclusions Infrastructure failures may invalidate a pair; collisions, safety failures, timeouts, and task failures remain scientific outcomes.
Keep null results valid A near-zero causal effect is scientifically acceptable. A broken pairing mechanism is not.

04 / SYSTEM PANEL

Compare intervention architectures without claiming them as novelty.

CONTROL π₀.₅ + SafeLIBERO No defense. Establish nominal behavior and reproduction.
REACTIVE AEGIS / VLSA Post-hoc CBF/QP action projection. Primary intervention mechanism.
ANTICIPATORY / HYBRID Critic-VLA Public comparison arm after the AEGIS causal pilot passes.

The novelty claim is not “VLA safety exists” or “CBF projection works.” The candidate contribution is paired counterfactual measurement of how safety interventions change later nominal VLA proposals, plus architecture-level comparison under matched conditions.

05 / CURRENT FRONTIER

No scientific result is claimed yet.

✓
SafeLIBERO simulatorvalidated
✓
Policy environment + checkpointvalidated
→
π₀.₅ model restore + live querycurrent frontier
○
Paired Branch N / Branch S pilotnot completed

The current execution phase is infrastructure bring-up on Kaggle T4 x2. The simulator path, environment, checkpoint download, fixed-state restore, and rendering path are validated. The next gate is a live π₀.₅ policy query, followed by reproducible paired-noise instrumentation and only then the causal intervention pilot.

06 / WHY IT MATTERS

Safety layers can create a second-order control problem.

A safety correction may solve the immediate risk while changing the trajectory the learned controller subsequently sees. If that altered state systematically changes later nominal proposals—or increases future intervention burden—then evaluating a shield only on immediate collision avoidance misses part of the closed-loop behavior.

The project is designed to separate that causal question from ordinary benchmark performance, generic failure recovery, and stale-action-chunk deployment effects.