- CRISP // TUM 2025
- 1 KHZ CONTROL LOOP
- MULTI-ROBOT
CHIRON
THE BRIDGE BETWEEN MIND AND BODY
VLA models predict what a robot should do. But translating neural network outputs into safe, real-time physical commands at 1 kHz — with impedance control, workspace boundaries, and emergency stops — is an unsolved deployment problem. CHIRON closes the loop between AI policy and real robot hardware.
MODULE STATUS: ACTIVEControl Loop Frequency
1kHz
- DIVISION
- ANIMA
- WAVE
- W2
- DOMAIN
- ACTION
- WAVE 4 // ANIMA SUITE
- ROS2 VLA CONTROLLERS
VLA OUTPUTS DON'T SPEAK ROBOT
Vision-language-action models predict what a robot should do, but translating predictions into physical robot commands requires careful control engineering. Most VLA deployments stop at the neural network output — they don't address the hard problem: how do you safely close the loop on a real robot at 1 kHz?
Real robots need impedance-based compliance (not just rigid control), workspace boundary enforcement, emergency stops, and gravity compensation. There's no standard bridge between VLA policy outputs and industrial robot control interfaces. Every lab reinvents this stack from scratch.
WHAT CHIRON DELIVERS
CHIRON implements the CRISP framework (Compliant ROS2 Controllers for Learning-Based Manipulation Policies) — a production-ready control layer that converts VLA action predictions into real-time robot commands.
CAPABILITIES
- 7D VLA actions (6D pose + gripper) → real-time joint commands at 1 kHz
- Cartesian impedance, operational space, joint impedance, hybrid force-position control
- Real-time safety: velocity/torque monitoring, workspace boundaries, emergency stops
- Multi-robot: Franka Panda/FR3, KUKA IIWA, Kinova Gen3 (extensible)
- C++ backbone (<1ms cycle) + Python FastAPI frontend + per-robot adapters
- WebSocket streaming at 100 Hz + ROS2 topic subscriptions for VLA actions
WHY THIS IS HARD
Real-time robot control at 1 kHz requires solving four coupled problems simultaneously:
- 01Real-time scheduling: C++ backbone with <1ms cycle time, Python integration without GIL blocking
- 02Hardware abstraction: support Franka libfranka, KUKA ROS2, Kinova gRPC with unified interface
- 03Safety first: monitor velocity, torque, and workspace at every control cycle — stop immediately on breach
- 04VLA integration: convert normalized 7D actions to robot-specific joint commands with rate limiting and saturation
- 05Impedance tuning: stiffness 0-1000 N/m per axis, damping ratio 0.5-2.0 — critically damped default
CHIRON achieves this with a C++ controller backbone (1 kHz loop), Python FastAPI frontend, and per-robot adapter modules. Safety is non-negotiable.
REAL-TIME PERFORMANCE
Hard real-time metrics on production hardware:
| METRIC | VALUE |
|---|---|
| Control Loop Frequency | 1 kHz (1ms cycle) |
| Control Loop Latency | <1 ms |
| State Query Latency | 0.5–1 ms |
| Command Processing | 0.2–0.5 ms |
| WebSocket Update Rate | 100 Hz |
| Concurrent Connections | 100+ supported |
| Supported Robots | Franka, KUKA, Kinova |
| Control Modes | 4 (impedance, op-space, joint, hybrid) |
| Stiffness Range (Cartesian) | 0–1000 N/m per axis |
| Damping Ratio | 0.5–2.0 (critically damped) |
| Max Velocity Monitored | Configurable per joint |
| Max Torque Monitored | Configurable per joint |
| Workspace Boundary | Configurable min/max XYZ |
| Emergency Stop Latency | <2 ms |
WHAT'S BUILT TODAY
| COMPONENT | STATUS | NOTES |
|---|---|---|
| Core models | COMPLETE | CRISP controller stack validated — production-ready |
| Cartesian impedance | COMPLETE | Stiffness + damping in task space with gravity comp |
| Operational space | COMPLETE | Task-space with null-space optimization |
| Joint impedance | COMPLETE | Joint-space PD with gravity compensation |
| Hybrid force-position | COMPLETE | Contact-aware manipulation (F/T monitoring) |
| 1 kHz control loop | COMPLETE | C++ backbone, predictable <1ms latency |
| VLABridge | COMPLETE | Action scaling, rate limiting, safety filtering |
| SafetyMonitor | COMPLETE | Real-time velocity/torque/workspace enforcement |
| FastAPI REST API | COMPLETE | /command, /state, /configure, /emergency_stop |
| WebSocket streaming | COMPLETE | Real-time robot state at 100 Hz |
| ROS2 integration | COMPLETE | Topic subscriptions for VLA actions |
| Docker deployment | COMPLETE | Multi-container with Prometheus/Grafana |
| Franka support | COMPLETE | Via libfranka integration |
| KUKA support | COMPLETE | Via ROS2 LBRCIF interface |
| Kinova support | COMPLETE | Via gRPC API |
| API layer | IN PROGRESS | Public REST + gRPC API — pending dataset infrastructure |
| Kubernetes | PLANNED | Manifests in k8s/ directory |
WHERE CHIRON DEPLOYS
- APP_01
MANIPULATION LABS
Standard VLA deployment layer for research — CRISP-compliant baseline for learning-based control experiments.
- APP_02
FACTORY AUTOMATION
Safety-certified ROS2 control for collaborative robot arms on production lines.
- APP_03
LOGISTICS & WAREHOUSING
Reliable impedance control for pick-and-place operations at warehouse scale.
- APP_04
COBOT MANUFACTURERS
Drop-in integration layer between AI policy engines and robot hardware platforms.
- APP_05
RESEARCH INSTITUTIONS
CRISP-compliant baseline controller for benchmarking learning-based manipulation policies.
- APP_06
MULTI-ROBOT FLEETS
Unified VLA controller interface across heterogeneous robot hardware — Franka, KUKA, Kinova.
UNDER THE HOOD
FOUNDATION: CRISP (TUM, 2025)
- Real-time C++ controller backbone (libfranka, ROS2, custom)
- 1 kHz control loop with <1ms latency guarantee
- Gravity compensation via inverse dynamics
- Cartesian impedance: compliance in task space for contact-rich tasks
- Operational space: redundancy resolution with null-space optimization
- Hybrid force-position: simultaneous force and position control during contact
KEY INNOVATION
CHIRON bridges the gap between neural network policy outputs and real-time robot control — converting normalized VLA actions into safe, impedance-controlled joint commands at 1 kHz across multiple robot platforms.
DEPLOYMENT STACK
- C++ controller backbone + Python FastAPI frontend
- Per-robot adapter modules (Franka, KUKA, Kinova)
- Prometheus + Grafana monitoring
- Docker multi-container with ROS2 bridge
ROBOT INTERFACES
- Franka
- libfranka — real-time control at 1 kHz, impedance + force sensing
- KUKA
- ROS2 LBRCIF — interpolated control, industrial-grade reliability
- Kinova
- gRPC API — goal-based control, lightweight integration
RESEARCH PAPER
- [01]CRISP: Compliant ROS2 Controllers for Learning-Based Manipulation Policies — TUM, September 2025