Simulation only (MuJoCo). No hardware.
A robot with a deliberately dumb, frozen brain, and a designer that may only change its body. Brain A drives both motors
full forward with no sensors. Brain B adds a single bumper reflex. An LLM designer (Claude through the subscription CLI)
edits a Python body script built on MuJoCo's mjSpec API: wheels, casters, bumpers, whiskers, tails, springs, ballast.
Failure logs tell it where each attempt got stuck, flipped or slowed. Random search and CMA-ES search the same design space
with the same budget; the LLM only counts as having shown something if it beats them.
Questions:
- How much task competence can a smart designer move into passive mechanics for a brain that never changes?
- Does an LLM designer beat random search and CMA-ES at equal budget?
- Given different worlds (sand, ice, stairs, maze, steep hill, narrow corridor), do separately evolved bodies diverge? ("Darwin's finches".)
Plan, design-space schema, scoring and budget accounting: PLAN.md. Objections we have to survive: CRITIQUES.md. Third-party components: THIRD_PARTY.md.
M0 (body builder, Brain A, course 1, validity gate, scoring, tests) and M1 machinery (course 2, random search, CMA-ES, the LLM design loop, held-out report). Results are kept outside git; this README does not quote any.
dumbrobot/ catalog, schema, bodylib, body_template, worlds, validity, brains, sim, scoring (M0)
runlib, baselines, designer, design_loop, safeio, render, report, m1 (M1)
scripts/ run_llm_batch.sh (resumable `claude -p` batches)
tests/ pytest suite (the LLM loop is tested with a fake designer; no test calls `claude`)
PLAN.md CRITIQUES.md THIRD_PARTY.md
All arms get the same budget of proposals (invalid ones count), the same schema, validity gate, scorer and courses (bumps, corridor); winners are re-evaluated on held-out start-heading seeds.
python -m dumbrobot.m1 run --arm random|cmaes --seed K # baselines (CPU, via ~/projects/bin/sim-run)
python -m dumbrobot.m1 probe # one isolated `claude -p` call, emails redacted
scripts/run_llm_batch.sh llm_param 0 1 2 3 4 # LLM designer runs, resumable
python -m dumbrobot.m1 report # held-out re-evaluation, tables, rank tests
Every run is resumable: proposals and designer replies are cached on disk before they are spent.
python -m pytest -q
CPU simulation on the shared machine goes through ~/projects/bin/sim-run (one thread, nice 10) with total load
kept under 20. Evaluate the hand-made example with:
python -m dumbrobot.cli score examples/hand_made_big_wheels.py --course bumps
Apache-2.0, see LICENSE.