xhail/uav-array-codesign-rl

uav-array-codesign

An RL environment for coding agents, in Harbor / Terminal-Bench-Science format. The agent designs a 2x2 antenna array for a 2.45 GHz drone radio, then writes a beamforming and link controller that keeps the link up while the drone flies and gets jammed. A physics simulator scores the result, and the reward is 1 only if all 6 tests pass.

How it's graded

• Simulator files were not modified
• Antenna design meets the matching, isolation, gain and scan targets
• Submitted antenna characterization matches the simulator
• Controller holds the link on 5 public flight missions
• Submitted evaluation summary matches a fresh run
• Controller holds the link on 5 hidden missions

Exact thresholds are in the task README.

Built with Beam

We ran Reflection's Beam-1.0 as the agent through Harbor's terminus-2 harness, served by Hugging Face Inference Providers. Its antenna design passed every antenna test, but its controller dropped the link too often: 14 outages on the public missions (limit 10) and 15 on the hidden ones (limit 12). It saw the 14 outages and still declared success. Trajectories and verifier logs are in jobs/beam-1.0-terminus2/.

Run	Tests passed	Reward	Outages (public / hidden)
Oracle solution	6/6	1.0	3 / 3
Independent solution	6/6	1.0	8 / 7
Beam¹	4/6	0.0	14 / 15
DeepSeek V4.1 Flash	3/6	0.0	27 / 19
Starting baseline	1/6	0.0	584 / 551

¹ Ran under the earlier 15-minute agent budget. Only completed Beam trials are shown.

Quick start

T=tasks/engineering-sciences/electrical-engineering/uav-array-codesign
harbor trials start -p $T -a oracle

# Beam via Hugging Face (token limits below are a conservative guess)
OPENAI_API_KEY="$HF_TOKEN" OPENAI_API_BASE=https://router.huggingface.co/v1 \
harbor run -p $T -a terminus-2 -m openai/reflection/Beam-1.0-501B-A23B:baseten \
  --ak record_terminal_session=false \
  --ak interleaved_thinking=true \
  --ak 'llm_call_kwargs={"extra_body":{"chat_template_kwargs":{"allow_missing_thinking":true}}}' \
  --ak 'model_info={"max_input_tokens":131072,"max_output_tokens":32768,"input_cost_per_token":0,"output_cost_per_token":0}'

Layout

tasks/   the Harbor task: instruction, environment, solution, tests
task/    working copy, author notes and qualification script
jobs/    recorded agent runs (Beam, DeepSeek)