efa-flow-arm3
EFA (Energy First Architecture) open-weight release #2 — the flow-matching actuation policy for a 3-joint coupled
chain (6-D state, 3 torques): a velocity field v(s,a,t) integrated K forward passes from a=0. The corrected
2026 recipe — no iterative energy descent over actions, no BPTT — on the body where the discrete Gᵈ approach degrades.
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
Why this model matters
The eval-budget scaling result, as weights: a discrete argmin controller on this body costs 125–152 action-evaluations
per decision (exponential in DOF) and, in the flagship run, reaches 57%; this flow policy reaches ~100% at a
single forward pass (constant in DOF) — cheaper and better as DOF grows. The "student beats teacher" effect
(distillation exceeding its demonstrator's closed loop) was confirmed across three teacher variants in the ledger.
Architecture & inference
- Velocity field
v(s,a,t): relu/linear MLP over [cos(θᵢ−gᵢ), sin(θᵢ−gᵢ), ωᵢ for i=1..3, sin θ₁..₃, a₁, a₂, a₃, t] → 3-D velocity.
- Inference (K=1 sufficient):
a←0; for k in 0..K { t=k/K; a += v(s,a,t)/K } — clamp to ±UMAX and apply.
- Trained by conditional flow matching to a two-stage argmin demonstrator over a fitted value (HV=128).
Metrics (this artifact — gated, round-trip-verified; exact numbers in config.json)
Release gate: reach ≥95% at K=1, best-of-≤3 seeds; shipped weights reloaded from this file and re-evaluated
before release. Context from the ledger: flagship run flow 100% @ K=1 vs teacher 57% @ 152 evals; multi-seed study:
2/3 lightened-recipe seeds reached 100%, one seed's failed value-training capped its flow at 25% — the variance source
is FVI value-training stability, and the gate exists precisely because of it.
Honest limits (read before using)
- Simulated toy body (3-link chain, dynamics in
config.json); fully-actuated — the underactuated variant is a
measured open boundary (ledger: both teacher and flow ~22%; a greedy 1-step value has no plan for pumping a passive joint).
- Distilled: quality is bounded below by nothing but bounded in kind by the demonstrator's coverage; the flow
exceeds its teacher's closed-loop but cannot exceed its plan.
- Not energy-first per se: this is the plain velocity net (the hybrid sibling,
efa-hybrid-arm2, carries the
potential); a gradient-structured 3-DOF hybrid is future work.
- Trained & evaluated entirely on-device via Ferric; no external data.
Provenance
Built by experiments/efa_release3.rs (train → gate → save → reload → re-verify). Program record: WHITEPAPER.md
§3.9, docs/RESULTS.md §VII, docs/FRONTIER-CHECK-2026.md (the 2026 frontier check that corrected the recipe).
License: Apache-2.0.
