Robots That Feel: Safe Physical Collaboration with Humans

Author: Denis Avetisyan

Researchers have developed a new control system enabling legged robots to safely and compliantly interact with people during complex manipulation tasks.

A unified model and learning-based control system facilitates secure human-quadruped robot collaboration-complete with a manipulator arm-demonstrated effectively across both simulated environments and physical hardware.

A combined model- and learning-based controller with force control and a Reference Governor guarantees safety during whole-body loco-manipulation.

Coordinating complex whole-body movements with robust safety guarantees remains a key challenge for legged robots operating in dynamic environments. This is addressed in ‘Safe Whole-Body Loco-Manipulation via Combined Model and Learning-based Control’, which introduces a novel controller combining model-based admittance control for manipulation with a reinforcement learning policy for locomotion. The resulting system enables compliant and safe interaction through unified 6-DoF force response, validated via simulation and hardware experiments on a quadrupedal robot. Could this hybrid approach unlock more intuitive and reliable human-robot collaboration in increasingly complex real-world scenarios?

Deconstructing Control: The Limits of Precision

Conventional robotic control systems frequently encounter difficulties when deployed in real-world settings due to an inherent reliance on precise, pre-defined parameters. These systems are typically engineered for highly structured environments and specific tasks, exhibiting diminished performance when confronted with unexpected obstacles, variations in terrain, or novel situations. The rigidity of these control architectures stems from the challenges in accurately modeling and predicting the complexities of dynamic environments, leading to instability or failure when faced with even minor deviations from anticipated conditions. Consequently, robots operating in unpredictable scenarios – such as search and rescue, disaster response, or even navigating a cluttered home – often require constant human intervention or exhibit limited autonomy, hindering their potential for truly versatile and independent operation.

A fundamental hurdle in robotics lies in the intricate dance between movement and action; versatile robots require seamless coordination of locomotion – how they move through space – and manipulation – how they interact with objects. Current systems often treat these as separate problems, leading to jerky, inefficient, or even unstable behavior when attempting complex tasks. The difficulty arises because coordinating these functions demands constant recalculation of balance, trajectory, and force application, all while accounting for unpredictable external disturbances. Successfully integrating locomotion and manipulation isn’t simply about adding more powerful motors or sophisticated sensors; it necessitates developing control algorithms that anticipate, adapt to, and ultimately harmonize these two critical robotic capabilities, paving the way for robots that can fluidly navigate and interact with dynamic, real-world environments.

Robotics is shifting toward control strategies that prioritize immediate sensory input and dynamic adjustments over rigid, pre-programmed instructions. This move away from sequential actions necessitates algorithms capable of processing a continuous stream of data – visual, tactile, and proprioceptive – to understand the robot’s environment and its own state in real-time. Instead of executing a fixed series of movements, adaptable robots leverage this information to modify trajectories, adjust grip force, or even replan entire tasks on the fly. Such responsiveness isn’t merely about speed; it’s about creating a system that can gracefully handle unexpected disturbances, navigate uncertain terrain, and interact with objects in a nuanced and flexible manner, ultimately bringing robots closer to the levels of dexterity and improvisation seen in biological systems.

This framework integrates neural and AI methods with Model Based Design to enable safe and effective human-robot collaboration.

Orchestrated Movement: The Promise of Whole-Body Control

Whole-Body Control (WBC) represents a control architecture designed to simultaneously manage the degrees of freedom of a robot’s entire kinematic chain, encompassing both locomotion via legs and manipulation with arms. Unlike traditional approaches that often address locomotion and manipulation as separate tasks, WBC treats them as a unified problem, enabling coordinated movements and complex interactions with the environment. This is achieved by defining a centralized control objective – typically a desired end-effector pose or a target trajectory – and distributing the required torques and forces across all actuated joints. Consequently, WBC facilitates tasks requiring simultaneous leg stabilization and arm manipulation, such as maintaining balance while reaching for and manipulating objects, or navigating uneven terrain while performing dexterous tasks. The framework relies on accurately calculating the inverse kinematics and dynamics of the robot to achieve precise and stable control of its entire body.

Robust state estimation is critical for Whole-Body Control, providing the necessary data to accurately determine a robot’s configuration in space. This process involves fusing data from multiple sensors – including joint encoders, inertial measurement units (IMUs), and potentially vision systems – to create a consistent and reliable representation of the robot’s position, velocity, and orientation. Kalman Filters are frequently employed for this purpose due to their ability to optimally estimate the state of a dynamic system in the presence of noise and uncertainty; they recursively update the estimated state based on new measurements and a predictive model of the robot’s dynamics. The accuracy of the state estimate directly impacts the performance and stability of subsequent control algorithms, making a well-tuned state estimator a foundational component of the system.

Integrating Model-Based Control and Trajectory Optimization provides significant improvements in robotic performance during complex maneuvers. Model-Based Control utilizes a dynamic model of the robot to predict its behavior and calculate appropriate control actions, enabling precise tracking of desired states. Simultaneously, Trajectory Optimization algorithms generate dynamically feasible and efficient paths by minimizing a cost function – typically encompassing time, energy, and tracking error – subject to the robot’s kinematic and dynamic constraints. This combined approach allows for the computation of optimal control inputs that maximize precision, minimize execution time, and reduce energy consumption, particularly beneficial in scenarios requiring intricate movements or operation within limited spaces. The optimization process often incorporates constraints derived from the robot’s physical limitations, such as joint limits and maximum velocities, ensuring the planned trajectory remains achievable.

Safe operation of robots utilizing Whole-Body Control in real-world environments necessitates the implementation of protective mechanisms to address unpredictable external disturbances and modeling inaccuracies. These mechanisms commonly include collision detection systems, force/torque sensors for anomaly detection, and reactive control strategies capable of overriding planned trajectories when unexpected contact or excessive forces are encountered. Furthermore, incorporating safety-rated actuators and redundant sensors provides a fail-safe capability, minimizing the risk of damage to the robot, its surroundings, or personnel. The selection and integration of these protective layers must adhere to relevant safety standards, such as ISO 10218 and RIA 15.06, to guarantee a demonstrable level of operational safety.

The admittance controller effectively tracks commanded end-effector velocities in all six degrees of freedom based on 6-DoF force/torque input, as demonstrated by the close tracking of linear velocities in the [latex]x[/latex], [latex]y[/latex], and [latex]z[/latex] directions, as well as angular velocities around those axes.

Constrained Freedom: Guaranteeing Safety Through Control

Admittance control operates by relating end-effector forces to desired velocities, effectively modulating the robot’s response to external contact forces. This is achieved through the definition of an admittance function – a dynamic relationship expressed as [latex]F = M(v_d – v) + D(v_d – v) + K(x_d – x)[/latex], where F represents the applied force, x is position, v is velocity, and M, D, and K are the mass, damping, and stiffness parameters, respectively. By adjusting these parameters, the robot can yield to external forces rather than rigidly resisting them, thereby reducing impact forces and preventing damage to both the robot and its environment. This compliance is crucial for safe human-robot interaction and operation in unstructured environments, allowing for predictable and controlled contact behavior.

The Reference Governor (RG) functions as a safety filter by modifying the robot’s desired trajectories, or reference signals, before they are sent to the low-level controllers. This adjustment is based on the Maximal Output Admissible Set (MOAS), which defines the boundaries of safe operation for the system. The MOAS is a set of achievable outputs, considering the robot’s physical limits and environmental constraints. By comparing the proposed reference signal to the MOAS, the RG projects it onto the feasible region, ensuring that the resulting commands will not violate established safety limits regarding joint velocities, accelerations, or forces. This proactive constraint enforcement prevents the robot from attempting motions that could lead to collisions or damage, even in the presence of model uncertainties or external disturbances.

Model Predictive Control (MPC) integrates with admittance control and the Reference Governor (RG) to provide proactive risk mitigation. MPC utilizes a dynamic model of the robot and its environment to predict future states and optimize control actions over a finite time horizon. By incorporating the constraints defined by the RG – derived from the Maximal Output Admissible Set (MOAS) – MPC formulates an optimization problem that minimizes tracking error while simultaneously ensuring all system limitations are respected. This predictive capability allows the robot to anticipate potential constraint violations before they occur, enabling preemptive adjustments to trajectory and force profiles, thus enhancing safety and stability during interaction.

Experimental results demonstrate the precision and stability of the proposed control system during human-robot collaboration, as quantified by a Mean Squared Error (MSE) of less than or equal to 0.005 m²/s² for linear force tracking components and ≤ 0.029 rad²/s² for angular components. These values were obtained through repeated trials involving collaborative tasks and represent the deviation between the desired force trajectory and the robot’s actual force output. The low MSE indicates minimal error in force tracking, contributing to safer and more predictable interactions with human collaborators.

This trajectory illustrates how the Reference Governor effectively guides the system towards a safe state while satisfying operational constraints.

Beyond Automation: Towards Symbiotic Human-Robot Partnerships

Effective human-robot collaboration hinges on a robot’s ability to safely and intuitively respond to physical interactions, and this is fundamentally achieved through precise force control. Utilizing Force/Torque (F/T) sensors – devices that measure forces and torques applied to a robot’s end-effector – allows the robot to ‘feel’ its environment and adjust its actions accordingly. This capability goes beyond simply preventing collisions; it enables robots to perform tasks with humans, such as collaborative assembly or assisted rehabilitation, by ensuring compliant and predictable behavior. By interpreting F/T data, robots can regulate applied forces, maintain stable contact, and even anticipate human intent, creating a seamless and secure working relationship where the robot acts as a responsive and reliable partner, rather than a rigid machine.

The Unitree Go2 robot, featuring a highly capable D1 Servo Arm, is emerging as a valuable platform for developing and validating sophisticated control algorithms intended for human-robot collaboration. Its dynamic quadrupedal locomotion, combined with the arm’s six degrees of freedom, allows researchers to test force control strategies in complex, real-world scenarios-moving beyond simulations and into physically representative interactions. This combination facilitates investigation into how robots can respond to external forces, maintain stability during collaborative tasks, and adapt to unpredictable human movements. The Go2’s relatively small size and affordability further contribute to its appeal as a versatile testbed, enabling wider access to advanced robotics research and accelerating the development of safer, more intuitive human-robot systems.

Visual Whole-Body Controllers represent a significant leap in robotic adaptability by integrating visual perception directly into the robot’s control system. Rather than relying solely on pre-programmed movements or force feedback, these controllers utilize cameras and image processing to ‘see’ the environment and react accordingly. This allows the robot to dynamically adjust its posture and movements in response to unexpected obstacles, changes in human positioning, or variations in task requirements. By processing visual data, the controller can compute optimal body configurations to maintain balance, avoid collisions, and ensure smooth, natural interaction with humans, effectively bridging the gap between pre-planned actions and real-world complexity. This capability is particularly crucial in collaborative settings where robots must share workspaces and respond in real-time to the unpredictable movements of their human counterparts.

The convergence of sophisticated control algorithms and increasingly capable robotic hardware is poised to redefine the landscape of human-robot interaction, extending assistance beyond pre-programmed tasks and into dynamic, real-world scenarios. This collaborative potential isn’t limited to industrial settings; robots are becoming adept at assisting in healthcare, providing support for elderly individuals, and even collaborating on complex assembly or repair work. The ability for a robot to understand and respond to human intent, coupled with precise force control and adaptable movements, unlocks applications previously deemed impractical. Consequently, advancements in this synergy promise a future where robots aren’t simply tools, but genuine partners, augmenting human capabilities and improving quality of life across diverse fields.

The pursuit of safe human-robot interaction, as detailed in this work regarding loco-manipulation, inherently necessitates a willingness to probe system limitations. The research actively addresses force control and safety guarantees-a deliberate attempt to understand where the boundaries lie. This echoes Donald Davies’ sentiment: “A bug is the system confessing its design sins.” The paper doesn’t shy away from acknowledging potential weaknesses within the control architecture; instead, it embraces them through the Reference Governor, effectively treating identified limitations not as failures, but as critical feedback for refinement. This process of controlled exploration, of pushing against established constraints, is central to truly understanding and improving robotic systems.

Pushing the Boundaries

The presented work achieves a degree of loco-manipulation fidelity sufficient for demonstrable safety, but that safety is, fundamentally, a constraint-a declared boundary. The true test isn’t whether the system avoids instability, but how gracefully it approaches it. Future iterations must deliberately explore those boundaries, introducing perturbations not as threats to be deflected, but as data points to be dissected. The Reference Governor, while effective, represents a conservative estimate of system capability. Can learning algorithms, operating at the edge of stability-guided by carefully constructed reward functions-begin to expand that boundary, redefining what constitutes ‘safe’ operation?

Current state estimation techniques, however sophisticated, remain abstractions of a complex reality. Legged locomotion, coupled with dexterous manipulation, introduces compounded uncertainties. The system’s reliance on accurate models-and its inherent vulnerability to model mismatch-demands investigation into more robust, potentially model-agnostic control schemes. The question isn’t simply “can it be modeled?” but “how much accuracy is necessary?” Perhaps intentionally introducing controlled discrepancies – forcing the robot to adapt to imperfect information – will reveal more resilient strategies.

Ultimately, this work represents a step toward embodied intelligence, but intelligence isn’t about flawless execution. It’s about resourceful recovery. The next phase must focus not on preventing failure, but on understanding-and exploiting-the mechanics of recovery itself. Only then can these systems truly move beyond pre-programmed compliance and exhibit genuine adaptability.

Original article: https://arxiv.org/pdf/2603.02443.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

Deconstructing Control: The Limits of Precision

Orchestrated Movement: The Promise of Whole-Body Control

Constrained Freedom: Guaranteeing Safety Through Control

Beyond Automation: Towards Symbiotic Human-Robot Partnerships

Pushing the Boundaries

See also: