Larkfen

Larkfen is an experimental robotics training project built around one hard architectural rule: policies, robot dynamics, and flight controllers should not depend on a particular simulation engine.

The current portability example runs one shared quadcopter model and one shared velocity flight controller in two independent backends:

                  robotics/
          shared contracts, robot model,
              and flight controller
                    /     \
                   v       v
          pybullet_sim/   mujoco_sim/

Both backends run the same hover, forward-flight, and turning command sequence and write an MP4 recording plus JSON telemetry to their own artifacts/ directory.

docker build -f pybullet_sim/Dockerfile -t larkfen-pybullet-sim .
docker run --rm -v "${PWD}/pybullet_sim/artifacts:/artifacts" larkfen-pybullet-sim

docker build -f mujoco_sim/Dockerfile -t larkfen-mujoco-sim .
docker run --rm -v "${PWD}/mujoco_sim/artifacts:/artifacts" larkfen-mujoco-sim

• robotics/ owns simulator-independent contracts, physical parameters, actuator math, demonstration commands, and flight-control algorithms.
• pybullet_sim/ owns only PyBullet world construction, robot instantiation, state/force adaptation, rendering, and lifecycle.
• mujoco_sim/ owns the equivalent MuJoCo-specific implementation.
• AI policies and training will live outside all simulator directories and communicate only through the shared contracts.

The two engines are not expected to produce bit-identical results. Cross-backend tests compare whether they remain within the same useful response envelope.