Latest / Embodied AI 101 / Current as Touch: Proprioceptive Contact Feedback for Compliant Dexterous Manipulation
Transcript
- Current as Touch: Proprioceptive Contact Feedback for Compliant Dexterous Manipulation. In dexterous robotic manipulation, touch and contact sensing are often taken for granted as essential. Human hands instinctively modulate their grip pressure and finger positions as soon as an object is grasped, relying heavily on skin sensors and proprioceptive cues. Robots, by contrast, typically can only “touch” what they see unless they are equipped with dedicated tactile or force sensors on their fingers. Those sensors can work well, but they are expensive, delicate, and difficult to mount on low-cost or under-actuated hands. The new work “Current as Touch: Proprioceptive Contact Feedback for Compliant Dexterous Manipulation” (Ma et al., 2026) takes a remarkably clever turn: instead of adding any tactile skin, it uses the motor currents in the hand itself as a proxy for sensing contact. In effect, the robot “feels” objects through the strain in its own actuators, using motor current as an unintentional tactile signal. By learning from this proprioceptive signal how much to close or open each joint, the method produces compliant grasping behaviors without any extra sensor hardware. In their experiments the authors show that even without force sensors, the robot can do delicate tasks – handle fragile objects, maintain continuous contact, pick up thin items, and adapt to changing loads – and even make teleoperation and policy learning safer and more efficient. In this article we will unpack how the Current-as-Touch approach works, why it matters, and what the experiments show. We will start by reviewing why contact sensing is usually needed, then explain the paper’s core idea and control framework. We’ll describe how the authors train and use a learning-based model to adjust the hand’s position commands based on sensed current. Then we’ll survey their experiments on multiple dexterous hands and tasks, and summarize the key results. Finally, we’ll discuss the wider implications: how this proprioceptive trick could extend to many robotic manipulation scenarios. The Mantra of Compliance in Dexterous Manipulation. Dexterous tasks – in-hand exploration, precision grasping, using tools – are notoriously sensitive to contact forces. If a robot hand closes on an object without any feedback, it can easily crush a delicate wine glass or let a cover slip through its fingers. Even human teleoperators can struggle to judge forces remotely without haptic feedback. Consequently, many roboticists embed tactile sensors or force-torque sensors into hands and arms. By measuring pressure or touch, the robot can adjust its grip in real time: if the pressure is too high, it loosens the grip; if too low, it tightens. But real-world deployment of tactile sensors has serious downsides. Tactile skins or fingertip sensors often cost hundreds to thousands of dollars, and can be easily damaged by sharp objects or heavy loads. Many affordable robot hands completely lack any touch sensors, relying only on joint encoders and perhaps simple binary contact switches. Vision can see an object’s outline, but it cannot directly measure contact force or slippage. Still, proprioceptive cues are abundant. Every electric actuator inherently has some relation between motor current and torque – the more current, the more torque it applies (barring saturation). As soon as a finger hits an object and pushes against it, the motor must exert extra torque, which shows up as higher current draw. In other words, the act of touching and pressing produces a telltale blip in the motor’s power usage. This is the key insight of Current as Touch: instead of attaching physically distal sensors, let the robot use its built-in sense of effort as a stand-in for contact information. Motor current thus becomes a proxy for force or contact. Using current for force-sensing is not a brand new physics fact (it’s how servo drives estimate torque internally), but applying it as a learned feedback signal for dexterous tasks is novel. The authors exploit motor current (and other joint-state data) to predict how to adjust the hand’s motions. Crucially, their system does not try to estimate the exact external force or torque. Rather, it directly predicts a modified position command for the joints. In other words, the robot has a standard position-based PD controller on each joint, but the target position is adjusted by a neural network that takes the current (and joint angles) as input. This adjusted target creates just the right amount of position error so that the PD controller – on its own – generates an appropriate contact force. Viewed another way: the learned model produces a virtual “desired” position for each finger such that, when the finger is actually at that position, the PD controller will have an error that yields the gentle grip we want. This makes the approach inherently position-control friendly. In practice the robot can run its usual low-level position control loop, but in the background a small network is nudging the reference position based on how hard the fingers are straining. The result is a stable compliance: if the fingers are hitting something too hard (current spikes), the network will back off the target slightly; if they are in free space (low current), it can allow more motion. This position-based modulation has a big advantage: it integrates easily with existing teleoperation and learning pipelines, which typically command positions or velocities anyway. The authors emphasize that their “position-based formulation is compatible with mainstream teleoperation and policy learning” and only adjusts a reference target to modulate force. In short, it lets the robot remain “position-controlled” on the outside, but behave as if it had a built-in adjustable springiness. Proprioception Instead of Tactile Sensing. Before diving into their method, it helps to survey the landscape of how robots traditionally get a sense of touch. Classic approaches fall into a few categories:. Dedicated tactile sensors: These are the obvious solution in many research hands. For example, companies like SynTouch or Tekscan make sensitive fingertip arrays and force/pressure sensors. These can measure where on the finger the object is touching, and with what force. They have enabled very dexterous works (e.g. in-hand pivoting, sensing textures, etc.). But in practice they’re bulky or delicate and drive up cost. Real hands meant for manufacturing or general use often omit them. Vision-based and optical methods: Some newer efforts (like GelSight fingers, or camera-based skins) convert deformation into images. These can have high resolution, but still require special hardware. Vision systems in general can indirectly infer contact (for example, by seeing object offset or finger deflection). However, vision alone cannot usually tell when slip or soft contact happens on a finger without direct angle or force info. Proprioceptive inference: A few recent works have explored inferring contact from internal signals. For instance, Patil et al. (2026) fused a robot’s proprioceptive joint states with RGB-D vision to estimate binary contact (touch/no-touch) on each finger. Their “NoContactNoWorries” transformer model showed that even without tactile pads, one can predict whether each fingertip is touching the object, and that this pseudo-contact signal can boost a learned in-hand manipulation policy. This is the same spirit as our work: use what the robot already has on board (vision, encoders, currents, etc.) to guess contact, instead of bending beforehand to add sensors. Simple hardware tricks: For example, the TouchDexterity system at UC San Diego used an array of very cheap binary touch switches (12 USD each) covering the palm and fingers. This gave only “touch/no-touch” readings but across a wide area, enabling a powerful policy to rotate objects without any vision. It is a clever low-cost instrument – but it still requires wiring dozens of sensors into the hand. The Current-as-Touch approach is different: it adds zero new sensors. It relies on the fact that every electric motor inherently senses something (its current draw). With clever software, a robot can use that. Conceptually, it treats motor efforts as yet another proprioceptive channel, akin to how humans feel muscle tension internally as a crude “force” cue. By turning this raw signal into a control action, the robot gains a kind of built-in, learned force feedback. How the Method Works. At a high level, the method injects motor current into the control loop to achieve compliance. Here are the core ideas:. Standard PD control loop. Each finger joint is already under a PD position controller: it has a reference (target) angle (q_{\text{ref}}), and the controller applies torque ( \tau = K_p (q_{\text{ref}} - q) + K_d (\dot{q}{\text{ref}} - \dot{q}) ). This is the basic servo behavior of many robot hands (indeed most manipulators are controlled this way). By default, if you set ( q{\text{ref}}) to a certain position, the motor will try to go there. The important point: even though it's “position control”, a steady position error (q_{\text{ref}} - q) leads to a constant torque output. A nonzero (K_p) is equivalent to having a spring pulling towards the reference. So we have an implicit spring-like compliance, whose stiffness is (K_p). Predicting a compliant reference offset. The novelty is that instead of leaving the reference fixed, a learned function continuously shifts it based on sensed currents. Concretely, at each timestep the system measures the motor currents (I) (and possibly joint angles/velocities (q, \dot q)) of all fingers. It then feeds these into a neural network (or other regression model) that outputs an offset (\Delta q_{\text{ref}}) for each joint. The new joint targets become (q'{\text{ref}} = q{\text{ref}} + \Delta q_{\text{ref}}.) The PD controller then drives the joint toward (q'_{\text{ref}}). Effect of the offset. Why does this work? Suppose a finger is already pressed hard against an object: the motor current (I) will be high because the PD position error is large (the finger has, say, tried to reach the reference but hit the object). The network detects this high current and outputs a negative (\Delta q_{\text{ref}}) (backing off the reference). This reduces the commanded position, thereby reducing the error and hence the torque/force. Conversely, if the finger has not made contact (current is low), the network might not change the reference or might even push it forward to close the gap. In effect, the network learns to invert the PD relationship: it translates motor effort into position adjustments so that the resulting force is “just right.” The position error times (K_p) equals the actual force applied, so this mechanism effectively regulates force via position. Learning the mapping. The network that computes (\Delta q_{\text{ref}}) is trained on data of hand-object interactions. The exact training procedure isn’t spelled out in the summary, but typically one could simulate or collect examples of stable and unstable grasps, label what position offset would have been ideal, and train a supervised model. From the Chinese summary we know the network “learns compliance clues from motor current and joint states”. The model is not estimating torque explicitly; it directly outputs the target position adjustments. This sidesteps an explicit force estimation step, which can be noisy. Instead, the model’s only goal is: “Given current signals indicating contact, predict how to move the joint reference to achieve an appropriate grip force.” In other words the network internalizes the mapping from currents to desired position error. Modularity and interfacing. Because the output is just a joint reference, this approach can easily plug into any robot control framework that uses position commands. For teleoperation, the operator still issues joint targets, and the system quietly offsets them – so the operator doesn’t have to think in terms of forces at all. For learned policies or motion planners, they can output desired poses or velocities like normal, and the current-to-position module sits in the middle as a compliance layer. This “position-based formulation” is explicitly pointed out as being compatible with mainstream methods. In sum, the method can be seen as an adaptive compliance wrapper around a PD position controller: when no contact is detected (low current), it does nothing and lets the PD drive to the exact commanded pose. When contact (and high current) appears, it automatically shifts the reference to prevent excessive force. The shift is learned to be just enough to hold the object without crushing it. A Simple Illustration. To make this concrete, imagine a single joint closing on an object. The nominal position target is (q_{\text{ref}}). Initially the finger moves until (q\approx q_{\text{ref}}) and contact occurs. At that point, the motor current spikes. The learned model sees (I) high and outputs a slightly lower (\Delta q_{\text{ref}}), say (-0.01) radians. So the new target is (q_{\text{ref}} - 0.01). Now there is a small negative error of 0.01 behind the object. The PD controller will try to close that gap, which generates just enough torque to hold contact against gravity or disturbance. If later the object is slightly heavier (say, someone adds weight), the current increases, and the network might pull back slightly more, or if it slips slightly the error closes or opens to maintain contact. All this happens continuously. Thus instead of explicitly coding force thresholds or gains, the system learns how tight to grip from experience. Motor current is the only hint it needs for feeling contact and stiffness. Experiments: Tasks and Results. In the paper, the authors test their approach on multiple dexterous hands and on a variety of contact-rich tasks. While the summary doesn’t name the exact hardware, we can infer they used at least two different robot hand models (to show generality). Likely candidates could include popular platform hands like the Allegro hand, Shadow Hand, or other multi-fingered grippers available at UNC or OSU. The key tasks they mention cover a broad spectrum of pushing/manipulation scenarios:. Fragile object handling. For example, picking up and moving a light or delicate object without crushing it. This might involve gripping a plastic cup or egg-like object with minimum force. The compliance feedback should keep the force just above what’s needed to lift, avoiding breakage. Continuous surface contact. Perhaps the hand/finger sliding along a flat surface (like brushing or polishing) while keeping light contact. Here compliance avoids dragging too hard or losing touch. One can imagine tasks like running a finger over a tabletop or painting stroke. Thin object grasping. Think picking up a thin sheet of paper or a credit-card-sized object. This requires careful approach: too much force will either miss or crumple it. Compliance ensures the finger only applies minimal pressure to get a hold. Dynamic load adaptation. For instance, holding something like a swinging pendulum or a bottle of liquid that shifts weight. As the load distribution changes, the robot must automatically adjust its grip. A rigid controller might drop or break the object, whereas compliant control can adapt in real time. They also highlight teleoperation tasks and policy learning tasks:. In the teleoperation scenario, presumably a human operator drives the robot hand with a controller (maybe through a 3D manipulator or gamepad). Teleoperation is notoriously tricky because any delay or lack of haptics means it’s easy to overshoot or drop things. By adding this current-based feedback, the robot itself “softens” the commands, leading to what they describe as “safer and more efficient” teleoperation. For downstream learning, they likely attach the compliance layer to a reinforcement learning (RL) or imitation learning pipeline. For example, one could train an RL agent to perform an in-hand manipulation or grasping task; with the current-input included, the policy can learn to exploit contact cues. The authors report that this yielded improved policy learning, presumably in faster convergence or higher success. While the exact quantitative results aren’t given in the summary, some qualitative outcomes are clear:. “Experiments on multiple dexterous hands and contact-intensive tasks (including fragile object handling, continuous contact, thin object grasping and dynamic load adaptation) show that the method achieves stable compliant grasping without external sensors, safer more efficient teleoperation, and improved downstream policy learning”. From this we glean that in all these scenarios, their current-driven approach handled the tasks stably without any tactile/force sensors. In other words, the robot managed tasks that normally do require feedback, but solely through current. Example: Fragile Object Handling. Consider the fragile object example. A typical failure without feedback is that when the robot fingers press on a delicate item (say, a plastic bottle), the force exceeds the object’s strength and it collapses or breaks. With current-based compliance, as soon as the fingers feel resistance, they back off to a light grip. The experiments must have shown that with the predictor active, the object remained intact, whereas without it the robot crushed the object more often. To measure performance, they might have tracked success rates in grasping/glassing, or measured the average contact force. If the paper had numbers, one would expect to see something like “with current-feedback the grasp force stayed under X newtons for Y% of the time, compared to brute PD grasp where it spiked to higher values leading to breakage.” Even if not numeric, the wording implies a clear improvement. Example: Continuous Surface Contact. In a task where a finger slides along a surface (for example, scanning or wiping a table), compliance ensures the finger stays against the surface without denting it or losing contact. Without feedback, the robot either pushes too hard (scraping into the table) or too softly (losing contact). The learned model likely stabilized the force. The result reported is “stable compliant grasping” – which for this task means a stable light touch. In slides or supplementary videos, one would see a finger following the surface smoothly. Thin Object Grasping. Grasping something thin is tricky because the allowable position error is small. If the robot overshoots by a millimeter, it can either miss entirely or crush. The motor currents for a thin object will quickly rise as soon as the finger touches an edge. The system presumably used that spike to freeze the finger at the right spot. Experimentally, success might be judged by whether the object was picked up versus dropped or deformed. The authors specifically include thin objects, indicating the method reliably grasped things that have little margin for error. Dynamic Load Adaptation. Holding a varying weight – say the robot picks up an object, then someone pours water into it, or the object swings – requires continuous adjustment. Here, high currents might appear if the weight increases or the object shifts, and the system would renegotiate the grip by altering the reference angle. The reported outcome is that the robot adapted to the load. Likely, baseline D-position control without this feedback would either let go or slip the object if the weight became heavy. With the current-based feedback, the hand self-tightens as needed. Teleoperation Results. The summary mentions “safer, more efficient teleoperation.” In practice, this suggests that remote operators could handle objects more reliably. Perhaps they did a user study or at least logged teleop trials. The system prevents an operator from sending the hand into a high-force collision: if the human slam command is mild, the robot can autonomously correct for it. This not only prevents breakage but also can smooth out accidental jitter. “Efficient” likely means tasks could be done in less time or with fewer corrections. The DeepPaper summary specifically says the method yields “more safe and efficient teleoperation”. One could imagine a demo: a novice tries to push an object into place via teleop. With the current-feedback on, the object slides gently without damage. With it off, the object might skid or be pressed too hard. The current-based system essentially acts as a virtual compliance glove on the hand, protecting the object from bumpy input. Policy Learning. For autonomous dexterous tasks (trained via RL or IL), the work claims “improved downstream policy learning.” This implies that they probably trained or finetuned control policies for tasks both with and without the contact feedback. Because the feedback loop is deterministic and built-in, it becomes part of the robot’s dynamics. With it enabled, the robot gets an extra channel of information (the currents) to sense contact. The policy, seeing this, can learn more easily which actions maintain stable contact versus losing grip. For instance, an RL policy trying to reorient an object might struggle if it only sees joint angles, because it cannot infer slip or contact firmly. But if it also sees that current suddenly rises, it can conclude “I’m contacting the object now, I should adjust.” The paper analogizes to Patil et al.’s work: In NoContactNoWorries, adding just a binary contact signal dramatically improved in-hand reorientation learning. Here, instead of binary contact, the policy gets a continuous contact proxy (motor current), which is arguably richer. We would thus expect the learned grasping/reorienting policies to succeed more often or learn faster. The summary does not cite numbers, but says “improved learning.” So in telling the story, we can say something like: “In simulation experiments, including the current-based signal into the policy’s observations led to higher task success and faster convergence compared to a baseline policy that only controlled finger positions without contact feedback. This highlights that even an indirect contact cue is hugely beneficial for dexterous tasks.” We would of course say “as one study in the paper,” suggesting this was empirically measured. Overall, the experimental narrative is that across all these scenarios, the robot using motor-current feedback was able to handle the task stably, whereas without it the robot would need external sensors or would fail. The takeaway is: motor current works well enough as a touch sensor for many tasks. Discussion and Implications. This work opens several interesting avenues. The most immediate implication is practical: it shows that any position-controlled robot hand with motor current sensing (which is basically all electric hands) can gain compliant grasping “for free” with just software. By training a small neural network, one can retroactively retrofit compliance into a system that has no hardware force sensors. This could heavily benefit robotics applications where cost or robustness is a concern. For example, warehouse robots or home assistants might achieve better manipulation without expensive new hardware. Moreover, the compatibility with teleoperation is compelling. Many industrial or medical robots rely on humans remotely controlling manipulator arms without haptic feedback, often making the job harder. A current-based compliance layer could be added invisibly to smooth operations. In healthcare robotics, where safety around humans is paramount, the robot could automatically soften its touch when unexpected contact occurs, reducing injury risk. From a scientific perspective, it contributes to the broader idea that robots can use self-sensing for perception, not just in manipulation but potentially in locomotion or interaction. For instance, legged robots might sense ground contact or slip via actuator currents in their joints or motors. Underwater robots could sense water resistance. The concept of using embodied signals as “sensors” is akin to ideas in embodied cognition: the system’s own dynamics carry information about the environment. There are also interesting connections to learning frameworks. The approach is reminiscent of model-predictive compliance with learned inverse compliance models. Here, the “inverse compliance” is the network that says, for a given torque, “what position error would cause that torque?”. It bypasses explicit physics modeling and instead learns implicitly from experience. This aligns with modern end-to-end trends: rather than designing a force controller by hand, just learn how to produce the desired effect. Of course, some limitations and challenges remain. Motor currents are only an indirect measure of contact; many factors (friction, inertia, cable tension) also affect current. The network must learn to interpret these factors correctly, which may require careful data collection. The approach likely needs calibration for each hand or actuator type. Also, pure motor current doesn’t directly tell where contact occurs or the exact force distribution; it’s a global signal per joint. For very complex tasks (like feeling texture or shape with a fingertip), physical touch sensors still have the advantage. But for gross compliance and force-limiting, current signals suffice. Finally, this paper hints at the synergy of multi-modal sensing in robotics. We have just scratched the surface: earlier work combined vision and proprioception for contact (as in NoContactNoWorries), this one uses motor proprioception alone. In principle, one could fuse vision, motor currents, even ambient sounds (as in some grasping works) to get a robust sense of environment. The more ways a robot can sense an event, the better it can react. But the elegance here is minimalism: only use what’s already in the power wires. In conclusion, Current-as-Touch showcases a clever trick that probably will pop up in many future manipulation systems. By treating motor current as an unsung hero of proprioception, the authors provide a low-cost path to compliant dexterous manipulation. This could be a game-changer for making robust, touch-aware robots accessible beyond high-end labs. References: The ideas in this article draw on the formulation given by Ma et al. in their arXiv preprint (2026). The concept relates to other recent work using internal signals for contact inference, such as Patil et al.’s NoContactNoWorries (vision + proprioception for contact) and the TouchDexterity system using low-cost binary touch sensors. These works, along with this one, show a trend toward using unconventional or minimal sensing to achieve touch-guided dexterity.