Stable Core
The runtime surface stays deliberately small.
Planner -> PlanDecisionSkillCatalogplus YAML recovery metadataTool,RobotBackend,Ros2Bridge,Memoryinstruction -> plan -> execute -> observe -> memory -> resume/replan
Rust • ROS 2 • Skills • LLM • Sim-first
roboclaw-rs is a Rust robotics workspace that keeps the runtime core narrow and pushes design
exploration into comparable experiment suites. One agent loop, one skill catalog, one backend boundary.
Stable Core
Planner -> PlanDecisionSkillCatalog plus YAML recovery metadataTool, RobotBackend, Ros2Bridge, Memoryinstruction -> plan -> execute -> observe -> memory -> resume/replanWhy it matters
Promotion, rollback, provenance, prompt shaping, recovery routing, observation reducers, and lag budgets
all compete in the experiments/ tree first. The runtime only adopts the minimum boundary that
survives repeated comparison.
Experiment → convergence
Compare multiple planner shapes against the same instruction set and metrics.
catalog score · pipeline router · keyword rulesStress retry, resume, and recovery behavior without mutating the stable runtime seam.
resume-aware replan · fail fast · direct resume firstModel how promotion, rollback, lag, and artifact trust should behave across release surfaces.
promotion rules · rollback budgets · stitching budgetsRuntime demo
$ cargo run --example pick_and_place -- \
"Use the simulator to pick up the red cube and place it in bin_a."
planner_provider=local
selected_skill=pick_and_place
step_completed=detect_object
step_completed=move_to_object
step_completed=grasp
step_completed=place_object
completed=true
Mock, local, and cloud planners target the same YAML skill catalog through one contract.
Each step can declare expectations, retry budgets, and resume checkpoints.
/cmd_vel, /joint_states, /roboclaw/action, and /roboclaw/state stay visible as the integration boundary.
Reference