Hugging Face's logo

Datasets:
xhail
/
uav-array-codesign-rl

RL Environment
Tasks:
Reinforcement Learning
Size:
n<1K
Tags:
reinforcement-learning
harbor
terminal-bench
beam
reflection
Libraries:
Harbor
License:
Dataset card Files
xet
 Community
uav-array-codesign-rl / task /Piyush.dev /BRIEF.md
xhail's picture
xhail
Publish uav-array-codesign Harbor RL environment with Reflection Beam evaluation
177fdd5 verified
preview code
|
Raw History Blame Contribute Delete
7.22 kB

Task brief: drone-hfss-link-codesign

Status: APPROVED Grader version: v2.0 Owner: Piyush Date: 2026-10-25

Capability claim

A pass proves that an agent can jointly solve a coupled electromagnetic structure design problem (3D array geometry, mutual-coupling active impedance, and L-section reactive matching network) and a sequential 6-DoF flight beamforming and link-control problem (body-frame attitude transformations, jammer nulling under active reflection constraints, and thermal/energy/queue management) from physical simulation telemetry. A fail does not distinguish between an electromagnetic mismatch failure and a flight-control failure without inspecting the per-test CTRF breakdown.

Seed

Inspired by real UAV command-and-control (C2) and telemetry link failures during high-bank maneuvers in contested RF environments:

  1. A belly-mounted 2x2 planar dipole array on a conductive carbon-fiber quadrotor airframe exhibits severe scan blindness and high return loss when designed with textbook free-space half-wave dimensions ($L = \lambda_0/2$) and an improper airframe standoff ($h \approx \lambda_0/2$), because finite wire radius adds inductive self-reactance, mutual coupling shifts the active impedance away from $50,\Omega$, and the carbon-fiber ground image cancels nadir radiation.
  2. In 6-DoF flight, steering the phased array using world-frame angles without transforming through the drone's instantaneous roll-pitch-yaw rotation matrix mispoints both the GCS main lobe and the spatial jammer null whenever the aircraft banks.
  3. Aggressive unregularized zero-forcing null steering excites anti-phase currents between coupled array elements, spiking the Total Active Reflection Coefficient (TARC), folding back power into the PA, and triggering thermal compression or shutdown.

Because the simulation engine (/app/hfss_sim/) couples Carter's Induced-EMF mutual impedance integrals, image theory, L-network transformation, 3D ZYX Euler kinematics, two-ray ground reflection over lossy soil, Doppler inter-carrier interference, and stateful PA thermal/queue/battery dynamics, no pre-trained lookup or textbook formula suffices without querying the physics engine and jointly optimizing the hardware and controller.

Proposal screen (7 criteria)

  1. Verifiable (Strong Accept): 100% deterministic analytical physics in NumPy/SciPy; zero LLM judges; verified in an isolated container (environment_mode = "separate", network_mode = "no-network").
  2. Well-specified (Accept): Exact file paths (/app/antenna_design.json, /app/hfss_characterization.json, /app/controller.py, /app/evaluation_summary.json), schemas, physical units, and target operating bands are stated.
  3. Solvable (Strong Accept): Solved via physical characterization of active array impedance + coordinate-descent/DE hardware design + body-frame active-TARC-regularized MVDR beamforming + feedback link control in <15 seconds of compute.
  4. Difficult for a good reason (Strong Accept): Requires multi-domain reasoning across electromagnetics (mutual coupling, image theory, L-network matching, TARC), 3D rigid-body kinematics (world-to-body frame transformation), array signal processing (regularized MVDR null steering), and sequential constrained control (PA junction temperature, packet backlog, RF energy reserve).
  5. Grounded and interesting (Strong Accept): Directly models the HFSS antenna-to-system workflow used in UAV RF link engineering.
  6. Scope (Strong Accept): Electromagnetics, RF systems engineering, and reinforcement learning / sequential control.
  7. Outcome-verified (Strong Accept): Verifier grades the physical performance of the submitted antenna design and closed-loop controller across both public and held-out 6-DoF flight scenarios.

Verdict: Strong Accept.

Why it is hard

  • Active vs. isolated impedance: Tuning the L-network to a single isolated dipole ($Z_{11}$) fails when all 4 elements radiate in phase because mutual impedances $Z_{12}, Z_{13}, Z_{14}$ and the airframe image elements shift the active feed impedance by $>35,\Omega$.
  • Airframe image physics: The conducting carbon-fiber belly acts as a reflector at distance $h$; naive $h = \lambda_0/2$ creates a broadside/nadir null via phase reversal on reflection.
  • Coupled TARC & MVDR nulling: Unregularized null steering toward a jammer close to the GCS direction excites high difference-mode currents, driving $\text{TARC}(\mathbf{w}) > 0.7$, which collapses mismatch efficiency and heats the PA past its $85^\circ\text{C}$ compression knee.
  • 3D Attitude rotation: The array is fixed to the drone body frame; world-frame line-of-sight vectors must be rotated by $\mathbf{R}_{\text{world}\leftarrow\text{body}}(\text{roll}, \text{pitch}, \text{yaw})^T$ at every 100 ms step.
  • Stateful resource dynamics: Transmit power and MCS selection must balance packet queue drain against Doppler ICI, PA junction temperature (cooled convectively by airspeed), and finite RF battery energy.

Expected expert time and target solve rate

Expert hours: 2.5 - 4.0 hours Target pass@k band for frontier models: 15% to 35%

Contract (what the agent is told)

  • Inspect /app/hfss_sim/ (the read-only electromagnetic & 6-DoF channel simulator) and /app/baseline/ (the failing legacy design and controller).
  • Produce:
    1. /app/antenna_design.json — physical 2x2 array geometry and L-network parameters.
    2. /app/hfss_characterization.json — S-parameter sweep and active scan-impedance report computed from hfss_sim for the submitted design.
    3. /app/controller.py — AdaptiveLinkController implementing reset(scenario_meta) and select_action(obs) -> dict.
    4. /app/evaluation_summary.json — closed-loop evaluation summary across the 5 public scenarios.

Acceptance (what proves done)

  1. /app/hfss_sim/ package files are unmodified.
  2. /app/antenna_design.json satisfies physical geometry bounds and achieves target return loss, impedance bandwidth, realized gain, and active scan reflection performance when evaluated by the independent verifier's HFSS engine.
  3. /app/hfss_characterization.json matches the independent HFSS engine recomputation for /app/antenna_design.json.
  4. /app/controller.py runs closed-loop with /app/antenna_design.json across all 5 public flight missions and meets all link reliability, delivery ratio, thermal, energy, and return thresholds, matching /app/evaluation_summary.json.
  5. /app/controller.py generalizes closed-loop with /app/antenna_design.json to 5 held-out adversarial 6-DoF flight missions in the verifier container without link loss, thermal overload, or energy exhaustion.

Scope

Agent may create/modify /app/antenna_design.json, /app/hfss_characterization.json, /app/controller.py, /app/evaluation_summary.json, and any helper scripts under /app/. /app/hfss_sim/ must not be modified.

Validation

  • scripts/lint-task.sh task/Piyush
  • harbor trials start --path task/Piyush --agent oracle (x2 determinism)
  • No-op, baseline, naive, alternative solution, semantic mutants (M1-M6), integrity/cheat mutants (I1-I4), per-instance null battery (Gate A).