physicalai-bmi/efa-2-acrobot

EFA-2 · Acrobot-v1 — the rescue artifact

The first EFA body where the energy gate doesn't just price compute — it rescues. Gym Acrobot-v1, exact published spec: underactuated (torque on the elbow only, actions {−1, 0, +1}), RK4 dynamics, reward −1/step, 500-step cap. Nothing about the task is ours.

Charlot Lab · Institute for Physical AI @ Bailey Military Institute. Runtime: Ferric. Siblings: efa-1 (flagship, certified multi-body) · efa-2-pendulum.

Why this body: the principled K=1 failure

From the hanging start, the optimal pumping torque is symmetric-bimodal (±1) — and a flow policy's conditional expectation at a₀ = 0 averages the modes toward zero. Measured on this artifact: K=1 rounds to zero torque on 54% of hanging-region probes. That is the classically multi-modal regime where one-shot policies fail — and it shows in the closed loop.

The rescue, measured (100 spec episodes each; never-solve = −500)

policy	mean return	solved	extra compute
flow K=1 (no gate)	−109.9	96%	—
AGENCY: E>τ → K=4 → planner	−88.0	100%	on 5.3% of decisions
flow K=4 always	−80.9	100%	on 100% of decisions (4×)
planner tool always (argmin own E)	−84.8	100%	3 E-evals every step
DP teacher (near-optimal)	−84.7	100%	106 evals/step

The model's own energy objects on ~5% of decisions, and that objection converts into solved episodes: +21.9 mean return, 96→100% solved, capturing most of always-K=4's gain at a small fraction of its marginal compute. The same potential verifies at 94.5% (ranks the demonstrator's action below the alternatives); every path — including both tools — is bit-exact deterministic. τ ships in config.json (95th pct of validation energy, from the artifact alone).

Honesty (read before citing)

• Distills a model-based DP demonstrator (grid value iteration on the known dynamics; teacher −84.7, 100%, squarely in the published good-policy band ≈ −80..−100). Not a model-free-RL comparison.
• K=1's failure is partial (96% solved but slow); the rescue Δ decomposes as ≈16 points from eliminating the four never-solves + the rest from faster solutions. One seed.
• ω observation components are scaled by their spec bounds (disclosed in config.json).
• Gated release: agency return ≥ −150 ∧ verify ≥ 90% ∧ bit-exact ∧ reload-exact — thresholds fixed before the run. Provenance: experiments/ebm_efa2acro.rs in the EFA repo (ledger with negatives included — on two easier bodies the gate found nothing to rescue, and the cards say so).

Architecture

Coordinated pair on the env's own observation [cosθ₁, sinθ₁, cosθ₂, sinθ₂, ω₁/4π, ω₂/9π]: flow head (8→96→96→1, CFM; K-step integration, rounded to the 3 legal actions) + potential head (7→96→96→1, contrastive; the verifier, the gate, and the planner tool's objective — one energy, three roles).

License: Apache-2.0.