When a human pushes open a heavy door in the dark, they do not need a pressure sensor on their fingertips to know how hard they are pushing. They feel the tension in their muscles and the resistance to their movement. In robotics, this internal awareness is called proprioception.
Traditionally, robots have relied on specialized hardware—such as multi-axis force-torque (F/T) sensors or tactile electronic skins—to measure interactions with the environment. However, these sensors introduce significant engineering challenges: they add weight, increase mechanical complexity, introduce high-frequency signal noise, and are prone to structural failure.
To circumvent these limitations, a powerful paradigm has emerged: sensorless force estimation. By leveraging the mathematical relationships between a robot's internal states—such as motor currents, joint positions, and inertial measurements—robots can accurately infer external contact forces entirely through software.
The Physics of Discrepancy: The Residual Observer
The fundamental principle behind sensorless force estimation is simple: any deviation between how a robot expects to move and how it actually moves must be caused by an external force.
A robot’s expected motion in free space is governed by its rigid-body dynamics model, which accounts for inertia, Coriolis forces, and gravity. If we know the commanded motor torques, we can predict the resulting joint accelerations. When an external force acts on the robot (such as a foot striking the ground or a hand grabbing an object), it introduces an additional torque ($\tau_{\text{ext}}$). The total torque equation expands to:
$$\tau_{\text{total}} = \tau_{\text{model}}(q, \dot{q}, \ddot{q}) + \tau_{\text{ext}}$$
By monitoring the "residual"—the difference between the expected torque from the dynamic model and the actual observed state—the robot can isolate $\tau_{\text{ext}}$ without any external sensors. For instance, Thomas Lew et al. demonstrate in Contact Inertial Odometry: Collisions are your Friends that monitoring this residual via an onboard Inertial Measurement Unit (IMU) allows a robot to detect and quantify impact forces during collisions, turning unexpected impacts into useful navigation data.
When explicit analytical models are difficult to derive—such as when handling complex, unknown objects—temporal histories of these deviations can be processed using machine learning. Haozhi Qi et al. leverage this in In-Hand Object Rotation via Rapid Motor Adaptation, where an in-hand manipulation system infers physical properties like object mass and friction by observing the history of joint movement deviations over time. Similarly, Linyi Huang et al. show in FunGrasp: Functional Grasping for Diverse Dexterous Hands that recurrent neural networks can map these tracking histories directly to external contact forces, enabling dexterous manipulation without electronic skins.
Motor Current as a Torque Proxy
How do we measure the actual torque ($\tau_{\text{total}}$) without joint torque sensors? The answer lies in the physics of electric actuators. For a standard electric motor, the output torque is directly proportional to the electrical current passing through its windings:
$$\tau = K_t I$$
where $K_t$ is the motor torque constant and $I$ is the current. By continuously monitoring motor current, the control software gains a direct window into the forces acting on the robot.
This current-based contact detection has been successfully deployed across diverse domains, from planetary exploration to collaborative manufacturing. The author of Results from InSight Robotic Arm Activities, E. Marteau, details how the InSight Mars lander's robotic arm achieved safe surface interactions by monitoring motor current fluctuations to detect contact in unpredictable extraterrestrial environments. On Earth, Zhongliang Jiang et al. utilize real-time motor current monitoring in Compliance control based on PSO algorithm to improve the feeling during physical human–robot interaction to allow collaborative robots to respond safely to human touch and manual guidance.
This proxy scales effectively to complex multi-legged systems. In Hur et al. demonstrate in Realization of Surfing Robot with Approximate ZMP Estimation Method without Force Sensor that motor currents can be used to calculate a humanoid robot's approximate Zero Moment Point (ZMP), enabling balance control on tilting platforms without foot force sensors. For smaller, cost-constrained platforms, the author of Corrigendum: Small-Sized Reconfigurable Quadruped Robot With Multiple Sensory Feedback for Studying Adaptive and Versatile Behaviors, Tao Sun, shows that monitoring current draw provides sufficient feedback for ground reaction force estimation and gait synchronization.
Impedance Control and Virtual Springs
Another elegant way to estimate force without sensors is to look through the lens of compliance and impedance control. Under a Cartesian impedance control framework, a robot is programmed to behave like a virtual spring-damper system attached to a target trajectory.
If the robot encounters an obstacle, it will deviate from its commanded position. Because the virtual spring constant (the stiffness matrix, $K$) is known, the physical force causing this deviation can be directly computed using Hooke's Law:
$$F = K \Delta x$$
This technique turns tracking errors into a high-fidelity "virtual force sensor." Guiyang Xin et al. exploit this property in Bounded haptic teleoperation of a quadruped robot's foot posture for sensing and manipulation, mapping the tracking errors of a quadruped's leg back to a human operator to provide realistic haptic feedback without contact sensors.
Similarly, the author of Towards Robotised Palpation for Cancer Detection through Online Tissue Viscoelastic Characterisation with a Collaborative Robotic Arm, Luca Beber, uses an impedance controller to reconstruct the resistance encountered during tissue palpation, effectively replacing physical force sensors with software-based observers to detect medical anomalies.
Filtering and State Estimation
While proprioceptive data is abundant, it is also notoriously noisy due to gear friction, backlash, and dynamic vibrations. To make these virtual force estimates useful for high-level tasks like walking or load adaptation, robots must fuse proprioceptive signals with probabilistic state estimators.
For legged locomotion, knowing precisely when a foot strikes the ground is critical for state estimation. Tzu-Yuan Lin et al. address this in Legged Robot State Estimation using Invariant Kalman Filtering and Learned Contact Events by combining an Invariant Extended Kalman Filter (InEKF) with a learned contact estimator that processes standard IMU and encoder data. To handle unexpected terrain variations, Mojtaba Hosseini et al. introduce a probabilistic "contact trust" model in State Estimation for Hybrid Locomotion of Driving-Stepping Quadrupeds, which dynamically weights the reliability of each limb's proprioceptive data based on its gait phase.
Furthermore, this model-based synthesis allows hardware-constrained humanoids to estimate complex global metrics. Grzegorz Ficht et al. demonstrate in Centroidal State Estimation and Control for Hardware-constrained Humanoid Robots that combining high-fidelity dynamic models with IMU data can synthesize center-of-pressure data, enabling advanced balance control on platforms completely lacking physical force-torque sensors. When a robot manipulates an unknown payload, Mateusz Jaszczuk et al. show in Rapid Mismatch Estimation via Neural Network Informed Variational Inference that internal joint torque data can be processed via variational inference to identify the inertial properties of a new load within milliseconds, ensuring compliant and safe human-robot collaboration.
Key Foundational Work
Realization of Surfing Robot with Approximate ZMP Estimation Method without Force Sensor (A Study on the Standing Stability of Robot Under Environmental Change in Direction of Gravity without Force Sensor) (2008) — contributed 2 concept elements to this direction.
2A1-N-065 Augmenting Robotic Sensory System Utilizing Its User's EMG and Posture Information(Cooperation between Human and Machine 3,Mega-Integration in Robotics and Mechatronics to Assist Our Daily Lives) (2005) — contributed 1 concept elements to this direction.