RFC 0002: Counterfactual-first metric design¶
- Status: Accepted for alpha
- Version: 0.1
Problem¶
Single recorded trajectories confound action effects with the initial scene and policy. A realistic future prediction may come from a scene prior while ignoring the candidate action.
Decision¶
The primary evaluation unit is an intervention group: one shared context paired with at least two uniquely named actions. Ground-truth futures are generated by restoring the same simulator state before each action branch.
For the LIBERO alpha, “same simulator state” includes MuJoCo mjSTATE_INTEGRATION plus
Python-side runtime state that can affect controls: robosuite clocks/controllers and
observable reset policy, global RNG state, and stateful gripper commands. The generator
must prove repeated-rollout and branch-order invariance; equality of LIBERO's flattened
time/qpos/qvel wrapper state alone is insufficient.
The alpha validates five complementary measurements:
- action dependence;
- counterfactual direction correctness;
- no-op stability;
- state trajectory error;
- candidate-selection regret.
Required sanity checks¶
Every new metric must document what it measures, what can game it, and how the following baselines should behave: exact oracle, copy-last, wrong-direction, and action-agnostic success prior. Learned metrics must additionally pin model weights and training data.
No composite score in v0.1¶
Action responsiveness, physical accuracy, and control utility are distinct. WAMProbe will publish a profile and paired confidence intervals before considering any aggregate score.