World-first unified MCP gateway + Control Barrier Function (CBF-QP) safety engine for agentic robotics. Gives 1000+ AI models (Claude, GPT-4o, Gemini) mathematically safe, real-time control over physical robots via ROS2.
14 tools with complete JSON schemas and descriptions, but significant quality gaps. Tool names lack clear action verbs (cbf_filter_velocity, predict_safe_trajectory use domain jargon; principia_ui, principia_status are vague). Descriptions are verbose (200-400 chars) and marketing-focused ('[WORLD-FIRST]') rather than LLM-optimized. Parameter descriptions exist but lack constraints (no enums, ranges, or format specs). Output schemas are documented in code but not formally declared in tool definitions. Error handling is minimal, no recovery guidance or actionable error messages. No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Obstacle parsing is implicit; array item schemas lack type definitions. Overall: functional but not production-grade.
[WORLD-FIRST] Parallel batch velocity filtering for fleet of N robots.
[WORLD-FIRST] Real-time CBF-QP safety filter for differential drive velocity commands. Takes proposed (v, omega) from AI and returns the minimally-perturbed safe command that guarantees h(x(t)) ≥ 0 ∀t (forward invariance).
[WORLD-FIRST] CBF altitude safety filter for quadrotors. Prevents drone from descending below z_min via real-time QP solve on thrust command.
[WORLD-FIRST] Combined CLF-CBF QP Controller. Simultaneously drives robot to target pose while enforcing hard barrier safety constraints.
[WORLD-FIRST] Dynamic CBF safety filter for moving obstacles with relative velocity vectors.
Full CBF safety audit: evaluates all barrier function margins for current robot state and returns per-obstacle safety analysis with actionable recommendations.
[WORLD-FIRST] Renders an ASCII safety radar grid showing surrounding obstacles and clear navigation paths.
Tool names lack clear action verbs. 'cbf_filter_velocity', 'predict_safe_trajectory', 'minimal_perturbation_proof' use domain jargon (CBF, QP) that LLMs may not map to intent. Rename to 'filter_velocity_for_safety', 'simulate_trajectory_with_safety_check', 'verify_command_optimality'.
Descriptions are verbose (200-400 chars) and marketing-focused ('[WORLD-FIRST]'). LLMs waste tokens parsing marketing copy. Trim to 50-150 chars, remove '[WORLD-FIRST]', state WHAT the tool does and WHEN to use it. E.g. 'Filter AI velocity commands to guarantee collision avoidance. Use before sending commands to hardware.'
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 38 | 2026-07-28+ | v2 |
[WORLD-FIRST] Evaluate real-time Lyapunov stability V(x)=½‖x-x_goal‖² and dV/dt along current dynamics. Checks if robot is mathematically converging to goal.
[WORLD-FIRST] Mathematically proves that CBF-QP modifies AI command with minimum L2 perturbation via KKT optimality conditions. Returns formal proof of optimality.
[WORLD-FIRST] Pre-simulate robot trajectory N seconds into future, evaluate all CBF safety constraints, and certify the command BEFORE any hardware executes it. Returns waypoints + safety certificate.
Run internal CBF-QP performance benchmark. Returns average solve time over 1000 iterations.
Get Principia Robotica server status, version, available tools, and mathematical capabilities summary.
Launch the local Principia Robotica interactive web UI visualizer and simulator in browser.
[WORLD-FIRST] Swarm fleet distance barrier checker. Evaluates inter-robot collision boundaries across N robots.
Array item schemas lack type definitions. 'obstacles' is declared as array but items are object with no properties schema. 'requests' in batch_cbf_filter has no item schema. This forces LLMs to guess structure. Add explicit item schemas with required fields and types.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). All tools are read-only but this is not declared. Agents cannot distinguish safe tools from destructive ones. Add 'readOnlyHint: true' to all 14 tools.
No error handling guidance. Tools return success/failure but do not tell LLMs what to do on error. E.g. if QP solver fails, should agent retry, adjust parameters, or ask user? Add recovery hints to error responses.