efa-hybrid-arm2

The first open-weight EFA (Energy First Architecture) model — the hybrid actuation architecture v = −κ∇ₐE + w for a coupled 2-link arm: one scalar potential Eθ(s,a,t) that actuates (descend its action-gradient) and verifies (Eθ(·,1): low = valid action), plus a small ℓ2-penalized correction net w that absorbs the residual a scalar fit cannot express.

Charlot Lab · Institute for Physical AI @ Bailey Military Institute. Runtime: Ferric (pure-Rust, cross-fabric: Metal / WebGPU / Vulkan / browser).

▶ Live demo — steer these exact weights in your browser (fetched from this repo, run on-device, no GPU): https://physicalai-bmi.org/assets/sims/efa-weights · WebGPU via Ferric-WASM: https://ferric.physicalai-bmi.org/efa

What this is (and is not)

This is a proof-point checkpoint, not a capability release: a tiny (~50k-param) goal-conditioned controller for a simulated 2-torque coupled pendulum chain, distilled from a fitted-value-iteration demonstrator. It exists to ship the EFA artifact class — the "coordinated family over one latent" as downloadable weights with a verified round-trip — and the release pipeline behind it. The architecture, not the body, is the point.

Architecture

• Potential Eθ(s,a,t): relu/linear MLP over [cos(θ₁−g₁), sin(θ₁−g₁), ω₁, cos(θ₂−g₂), sin(θ₂−g₂), ω₂, sin θ₁, sin θ₂, a₁, a₂, t] → scalar.
• Correction wφ(s,a,t): same inputs → 2-D velocity residual, trained jointly with an ℓ2 penalty (λ in config.json) so the potential must carry all the field it can.
• Inference (K=2 recommended): a←0; for k in 0..K { t=k/K; a += (−κ∇ₐEθ(s,a,t) + wφ(s,a,t))/K } — the action is energy descent plus a small correction. ∇ₐEθ is exact (analytic backprop through the relu net).
• Verify readout: rank candidate actions by Eθ(s, a, t=1) — lower is more valid.

Metrics (this artifact — gated, round-trip-verified; exact numbers in config.json)

Release gate: actuate ≥95% (K=2) AND verify ≥90%, best-of-≤3 seeds. The shipped weights were reloaded from this file and re-evaluated before release. Flagship-run context (see the validation ledger): actuate 100% @ K=1–4, verify 94.3%, potential carrying 65% of the field; λ-sweep frontier: ~57–65% energy-first at 100% control, knee between λ=0.1–0.3.

Honest limits (read before using)

• Simulated toy body (2-link chain, known dynamics in config.json); the policy is meaningless without that env.
• Distilled from a discrete FVI demonstrator — the win is the architecture and eval-budget (K forward passes vs Gᵈ argmin evals), not new capability.
• FVI seed variance is real (multi-seed study in the ledger); this artifact passed the gate, and the gate policy is disclosed in config.json.
• The potential carries the majority, not all, of the field — the measured energy-firstness is in config.json; a pure-potential variant costs actuation (ledger §VII).
• Trained & evaluated entirely on-device via Ferric; no external data.

Provenance

Built by experiments/efa_release.rs (train → gate → save → reload → re-verify) from the EFA validation program (65 measured experiments, negatives included): see WHITEPAPER.md §3.9 and docs/RESULTS.md §VII, including the 2026 frontier check that produced this architecture (docs/FRONTIER-CHECK-2026.md).

License: Apache-2.0.