Sub-Millisecond Reflexes for Continuous Physical Human-Robot Interaction

arXiv:2610.21913 · cs.RO, cs.SY, eess.SY · Submitted 2026-10-15 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: Today's paper: "Sub-Millisecond Reflexes for Continuous Physical Human-Robot Interaction".

Dev: The gist The proposed framework enables sub-millisecond reflex responses and uninterrupted physical human-robot interaction in continuous physical human-robot interaction tasks.

Rosa: First, who's behind it and why it matters.

Paper summary: Rosa: To summarize this paper, they are focusing on how we can achieve continuous physical human-robot interaction by having robots rapidly adapt their control behavior without interrupting what they are currently doing.

Dev: The core thesis is that current ROS two control architectures rely on stop-to-switch mechanisms which introduce significant delays and interrupt the task execution, but this work presents an opensource, sub-millisecond reflex engine built on an enhanced ROS two-compatible multimode controller <ref:2610.21913#pg1,ROS 2 control architectures rely on stop-to-switch mechanisms>.

Taro: They claim this engine enables deterministic switching between control modalities and integrates reflex behaviors specifically for mitigating contact hazards and kinematic limit violations before those limits are hit.

Rosa: The experimental validation is pretty impressive; they tested this on a Franka Robotics manipulator and showed switching delays below zero point one milliseconds, which is more than four orders of magnitude faster than conventional ros2 control, while maintaining continuous operation during safety-critical events <ref:2610.21913#pg1,more than four orders of magnitude faster than conventional ros2_control, while>.

Dev: They also showed that in a multi-stage manipulation task, the proposed framework reduced execution time by three point five percent and peak trajectory jerk by more than threefold compared to the standard ros2 control approach <ref:2610.21913#pg1,in a multi-stage manipulation task, the proposed framework reduced execution time>.

Taro: It’s interesting how they managed to smooth out those transitions too; they reduced peak operational-space jerk from four hundred ninety-nine point zero plus ninety-six point nine m/s3 down to one hundred fifty-five point four plus forty point six m/s3.

Rosa: So, the main point is that by using this framework, you get these reflex behaviors that help prevent rigid hardware stops and allow for seamless transitions between different control strategies during complex tasks.

Dev: And they achieve this by moving ROS two related instantiations and communication interface bottlenecks to the timing-insensitive initialization phase of the multimode controller itself <ref:2610.21913#pg1>.

Taro: It’s a clever way to handle that timing constraint without sacrificing the necessary responsiveness needed for real-time control loops, which is something I always look out for when testing these kinds of systems.

Conclusion: Rosa: So, to wrap up this paper, "Sub-Millisecond Reflexes for Continuous Physical Human-Robot Interaction," the authors have essentially presented a novel open-source methodology for safe, uninterrupted task execution without needing one single unified control strategy.

Dev: The implications are that we can combine the flexibility of hardware agnostic frameworks with the timing constraints and determinism typically associated with specialized systems, achieving switching delays below zero point one milliseconds <ref:2610.21913#pg1,switching delays below 0.1>.

Taro: This capability allows for reflex behaviors to execute within the one kHz control loop, enabling robots to react and recover during physical interaction before those rigid manufacturer safety stops are even triggered <ref:2610.21913#pg1>.

Rosa: They implemented five specific reflex behaviors—like Stop Reflex, Admittance Reflex, Joint Retraction, Cartesian Retraction, and Zero-G Reflex—to actively mitigate critical external forces and intercept joint position and velocity limits.

Dev: The framework demonstrated that a multi-stage task inspired by a real pHRI use case showed the benefits of seamless switching between control behaviors with different objectives, reducing task duration by three point five percent compared to standard ros2 control <ref:2610.21913#pg1>.

Taro: They also showed that when transitions are smoother, the maximum detected trajectory jerk is reduced by more than three times, which suggests better physical interaction for whatever the robot is doing in the real world.

Rosa: Ultimately, this work provides a foundation for learned robot policies where rapid transitions to established control strategies might improve robustness in unexpected situations during physical contact.

Jon Skerlj, Sami Haddadin

Munich Institute of Robotics and Machine Intelligence · Mohamed bin Zayed University of Artificial Intelligence

cs.RO, cs.SY, eess.SY

Submitted: 2026-10-15

Updated: 2026-10-15

Comments: Submitted to IEEE, source code available at: https://github.com/RoboRef/franka_reflex_engine

Code: https://github.com/RoboRef/franka_reflex_engine

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 85/100

The gist: The gist The proposed framework enables sub-millisecond reflex responses and uninterrupted physical human-robot interaction in continuous physical human-robot interaction tasks.

Key concepts

Multimode Controller (MC)
The MC acts as a central supervisor in the system, managing multiple lightweight subprocesses called controllets. It is responsible for orchestrating which specific control logic or task is currently active and ensuring resources are correctly assigned to each module at runtime.
Lock-Free Pointer Reassignment
To achieve extremely fast switching, the framework avoids traditional communication bottlenecks. Instead of complex data transfers, switching between control modes happens by simply reassigning a memory pointer within the active execution thread. This zero-overhead operation ensures transitions are instantaneous and deterministic.
Reflex Behaviours
These are pre-programmed reactive actions embedded in the controller that trigger when safety limits (like joint position or torque) are approached. Examples include Stop Reflex, Admittance Reflex, and Joint Retraction. They allow the robot to perform immediate, specific physical responses to external forces.
Controllet
A controllet is a modular, lightweight subprocess within the Multimode Controller framework. Each controllet is responsible for calculating the specific torque commands ($ au_{cmd}$) required for a single task or mode. They are assigned ownership of hardware resources at runtime to execute their designated control logic.

Summary

The gist The proposed framework enables sub-millisecond reflex responses and uninterrupted physical human-robot interaction in continuous physical human-robot interaction tasks.

How it works

  1. A real-time, sub-millisecond controller switching architecture is built upon the open-source multipanda ros2 framework, which provides native ROS 2 interfaces for coordinated control of multiple robots from a single process.

  2. The Multimode Controller (MC) acts as a central supervisor that orchestrates a modular suite of lightweight subprocesses called controllets.

  3. Each controllet is assigned ownership of one or more hardware resources in the ros2 control framework at runtime and implements the specific command torque τcmd calculation required for its assigned task.

  4. To achieve sub-millisecond determinism, all ROS 2-related instantiations and communication interface bottlenecks were relocated to the timing-insensitive initialization phase of the MC.

  5. Within the 1 kHz real-time execution thread, switching the active controllet is now reduced to a deterministic, lock-free pointer reassignment within the active memory space. This zero-overhead pointer swap successfully preserves the execution of the 1 kHz control cycle during transitions.

Hazard Management and Reflex Behaviours

The framework embeds 1 kHz reactive reflex behaviours into the multimode controller framework to prevent rigid hardware stops caused by manufacturer safety monitors.

The system evaluates joint-level position (q), velocity (q˙), and torque (τ) safety boundaries at each 1 kHz cycle. Safety boundaries are defined by checks for Position Limits (Hpos), Velocity Limits (Hvel), and Joint Torque Limits (Hτ).

The state machine transitions across four discrete states S ∈ NORMAL, REFLEX, WAIT FOR RECOVERY, RECOVERY.

Available reflex behaviours include five task-abandoning behaviours: Stop Reflex (freezes the manipulator), Admittance Reflex (maps external torques to reference joint velocities), Joint Retraction (displaces the target joint configuration along the negative joint torque vector), Cartesian Retraction (actively displaces the Cartesian target position by distance γ along the negative external contact force vector), and Zero-G Reflex (commands pure gravity compensation).

Experimental Validation

Empirical validation was conducted on a Franka Robotics manipulator.

Control behaviour switching delay was measured across 5 control behaviours for three distinct configurations: Baseline ros2 control, Original multimode controller, and Improved multimode controller.

The results showed that the baseline ros2 control resulted in approximately 210 milliseconds of delay, while the improved multimode controller ended with an average delay of 0.01 ms.

The switching latencies demonstrated that even the worst observed switching event of our proposed approach remained below 0.1 ms.

Task Performance Comparison

A multi-stage task execution was designed to evaluate real-world operational impact, involving three distinct parts: Pick & Place (fast velocity controller), Surface wiping (unified force-motion controller), and Tool Release & Return (stiff joint impedance controller).

The task performance of the enhanced multimode controller was benchmarked against the state-of-the-art ros2 control baseline.

The proposed approach reduced task duration by 3.5% compared with the standard ros2 control approach.

Smoother control behaviour switches utilized by the multimode controller reduced peak operational-space jerk from 499.0 ± 96.9 m/s3 to 155.4 ± 40.6 m/s3.

The framework demonstrated that the detected peak transient force (fmax) is significantly larger than the set collision threshold, and active feedback control engages well within Impact Phase II, eliminating static crushing forces before lower-level manufacturer safety limiters trip.

The paper presents a novel open-source sub-millisecond controller-switching methodology for safe, uninterrupted task execution without requiring a single unified control strategy. This approach combines the flexibility of hardware-agnostic frameworks with the timing constraints and determinism typically associated with specialized systems. It achieves switching delays below 0.1 ms, enabling reflex behaviours to execute within the 1 kHz control loop. Building on this capability, the framework implements five reflex behaviours to mitigate critical external forces and intercept joint position and velocity limits before rigid manufacturer safety stops are triggered. Furthermore, a multistage task inspired by a real pHRI use case demonstrated the benefits of seamless switching between control behaviours with different objectives. The proposed approach reduced task duration by 3.5% compared with the standard ros2 control approach. Smoother transitions also reduced the maximum detected trajectory jerk by more than threefold. Ultimately, our goal is not only faster controller switching but enabling robots to react, recover, and continue operating during physical interaction. The presented experiments are limited to a set of predefined reflex behaviours and safety thresholds. Future work will extend the library of available reflex behaviours and refine controller-specific parameters. The framework provides a promising foundation for learned robot policies where rapid transitions to established control strategies may improve robustness in unexpected situations.

--- Page 8 ---

The paper is supported by the Lighthouse Initiative Geriatronics by StMWi Bayern (Project X, grant no. IUK-1807-0007// IUK582/001) and LongLeif GaPa gGmbH (Project Y).

--- Page 4 ---

The paper was supported by the Lighthouse Initiative Geriatronics by StMWi Bayern (Project X, grant no. IUK-1807-0007// IUK582/001) and LongLeif GaPa gGmbH (Project Y).

--- Page 3 ---

The paper was supported by the Lighthouse Initiative Geriatronics by StMWi Bayern (Project X, grant no. IUK-1807-0007// IUK582/001) and LongLeif GaPa gGmbH (Project Y).

--- Page 2 ---

The paper was supported by the Lighthouse Initiative Geriatronics by StMWi Bayern (Project X, grant no. IUK-1807-0007// IUK582/001) and LongLeif GaPa gGmbH (Project Y).

--- Page 6 ---

The paper was supported by the Lighthouse Initiative Geriatronics by StMWi Bayern (Project X, grant no. IUK-1807-0007// IUK582/001) and LongLeif GaPa gGmbH (Project Y).

--- Page 5 ---

The paper was supported by the Lighthouse Initiative Geriatronics by StMWi Bayern (Project X, grant no. IUK-1807-0007// IUK582/001) and LongLeif GaPa gGmbH (Project Y)<ref:2610.

Improvements for AI systems

  1. Continuous physical human-robot interaction (pHRI) execution is enabled by enabling sub-millisecond reflex responses which allows robots to react to contact hazards, joint limits, and excessive velocities before manufacturer safety stops are triggered.

  2. The robot can execute a multi-stage manipulation task with reduced execution time by leveraging seamless transitions between control behaviours, eliminating the traditional stop-to-switch interruption, resulting in a demonstrated reduction of 3.5% in execution time.

  3. The system can perform compliant collision mitigation by integrating five specific reflex behaviours, including Admittance Reflex: Maps measured external torques τˆext to reference joint velocities via a proportional gain α, and Joint Retraction: Displaces the target joint configuration q∗d along the negative joint torque vector −τˆext scaled by Cartesian bounds δ = 0.05 m.

  4. The AI system can maintain continuous operation during safety-critical events by executing a reflex engine that intercepts hazards before reaching rigid hardware stops, as evidenced by a mean force mitigation time of 0.122 ± 0.013 s for external forces exceeding 15 N.

  5. The control architecture can achieve high determinism and low latency by implementing switching as a lock-free pointer operation within the real-time execution thread, eliminating runtime heap allocations and unnecessary data copying associated with standard ROS 2 middleware.

Related papers