Force Feedback and Humanoid Robotics
How Current Sensing and Adaptive Calibration Can Replace Expensive Tactile Skin at Scale
Force feedback in humanoid robotics is the ability of a robot to estimate and respond to contact forces — and for most humanoid tasks, this can be achieved by measuring actuator motor current rather than covering the robot in tactile skin.
One of the most persistent assumptions in humanoid robotics is that human-like touch requires human-like skin—dense arrays of capacitive or resistive tactile sensors covering every surface that may come into contact with the environment.
In research laboratories, this approach has produced impressive demonstrations. However, in the context of mass deployment, "Electronic Skin" creates a severe economic and reliability bottleneck. Tactile skin is expensive to manufacture, fragile under repeated contact, computationally heavy to process, and notoriously difficult to repair.
As humanoid robots move from prototypes toward mass production (e.g., Tesla Optimus, Unitree, Figure), a different engineering question becomes paramount: What is the minimum sensing required to perform a task safely, reliably, and cost-effectively?
For the vast majority of humanoid labor, the answer is not full-body tactile skin. It is Force Feedback derived from actuator current sensing, combined with Continuous Adaptive Calibration.
This article details the physics of current-based force estimation, the mathematical models required to compensate for environmental noise (temperature, humidity), and how high-fidelity linear actuators enable robots to "feel" without skin.
The cheapest sensor is the one you already have. If the actuator's mechanics are predictable, the current it draws already tells you what it's touching.
"Current-based force sensing only works when the actuator itself is mechanically predictable. A lead screw gives you a friction profile you can model in software; a harmonic drive gives you torque ripple you have to filter around. If you want a robot to feel through current, start by choosing an actuator whose mechanics don't fight the math." — Robbie Dickson, Founder and Chief Engineer of FIRGELLI Automations
1. Why is full tactile skin economically unviable for humanoid robots?
Covering a humanoid robot’s hands, fingers, arms, and torso with high-resolution tactile sensors introduces a cascade of compounding costs that render mass production unviable.
-
First-Order Costs (BOM): High-quality tactile arrays require exotic materials (piezoresistive fabrics), complex signal conditioning electronics (ADCs at every node), and protective coatings that must be both durable and sensitive.
-
Second-Order Costs (Integration): The wiring harness complexity explodes. Routing thousands of signal lines through rotating joints introduces failure points and increases the robot's base weight.
-
Third-Order Costs (Lifecycle): Tactile sensors degrade. Physical wear, hysteresis, and delamination mean that a fleet of robots would require constant, expensive skin replacements.
Most humanoid robots do not need to feel texture or temperature gradients on their forearms; they simply need to know if force is applied and how much.
2. How does an actuator measure force without a touch sensor?
In electric actuation, physics provides a powerful shortcut. We do not need a sensor to measure force if we can measure the energy consumed to create that force.
The Fundamental Actuator Equations
In a DC motor (the heart of most linear actuators), the relationship between current and force is governed by the Lorentz Force Law and mechanical transmission efficiency. (Source: Lorentz force law — Jackson, Classical Electrodynamics, 3rd ed.; motor torque-constant relationship τ = k_t · I is standard in Machinery's Handbook, motor and drive section.)
1. Current to Torque ($\tau$):
Where:
-
$\tau$ = Motor Torque (Nm)
-
$k_t$ = Motor Torque Constant (Nm/A)
-
$I$ = Current (Amps)
2. Torque to Linear Force ($F$):
For a linear actuator using a lead screw (like the FIRGELLI FA-BS16), torque is converted to linear force via the screw pitch:
Where:
-
$F$ = Linear Force (Newtons)
-
$\eta$ = Lead Screw Efficiency (0.0 to 1.0)
-
$L$ = Lead of the screw (meters per revolution)
3. The Combined Transfer Function:
Substituting the equations, we get the direct relationship between Current and Force:
The Insight:
The term inside the brackets is a constant (mostly). This means that by simply reading the current ($I$)—which acts as a proxy for torque—a robot can estimate the Force ($F$) exerted on the environment without a single external pressure sensor.
3. How does current sensing produce "virtual touch"?
When a robot equipped with current-sensing actuators (like the FIRGELLI FA-BS16) grips an object, the telemetry tells a story:
-
Free Motion: The current remains low and steady (overcoming only internal friction).
-
Contact: The finger hits the object. The motor slows down (velocity drops), but the PID controller tries to maintain speed.
-
The Spike: Current rises sharply to overcome the new resistance.
-
Force Estimation: The controller reads this current rise (e.g., +0.2A). Using the transfer function above, it calculates that 0.2A equals roughly 15N of grip force.
This allows the robot to detect contact, estimate grip strength, and prevent crushing—all using internal data.
4. What makes current-based force sensing unreliable in the real world?
If the world were perfect, the equation $F \propto I$ would be enough. It isn't.
The same force can produce different current readings depending on environmental conditions. To achieve "Expert Level" control, we must compensate for these variables.
A. Thermal Drift ($T$)
As the motor works, it heats up. Copper wire resistance ($R$) increases with temperature (Source: temperature coefficient of resistance for copper, α ≈ 0.00393/°C at 20°C — IEEE standard reference; Machinery's Handbook, electrical materials section):
While current control loops (Field-Oriented Control, FOC) handle resistance changes (Source: IEEE Power Electronics Letters literature on FOC; reference implementations in Texas Instruments InstaSPIN-FOC and Microchip motorBench documentation), temperature also affects the viscosity of the gearbox grease.
-
Cold Environment: Grease is thick $\rightarrow$ Higher friction $\rightarrow$ Higher current baseline.
-
Hot Environment: Grease thins $\rightarrow$ Lower friction $\rightarrow$ Lower current baseline.
-
Result: A robot calibrated in a warm lab might crush an egg in a cold warehouse because it underestimates the friction.
B. Altitude and Air Density
At higher altitudes, air density decreases, reducing convective cooling. Motors run hotter, shifting the thermal equilibrium point and altering the $k_t$ constant slightly due to magnet heating.
5. How does adaptive calibration solve environmental drift?
Instead of covering the robot in skin, we use software models to update the constants in our force equation.
The Adaptive Force Model:
Where:
-
$I_{load}$ = The current actually doing the pushing (what we want to measure).
-
$I_{friction}$ = Current wasted fighting grease/bearings (function of velocity $v$ and Temp $T$).
The Calibration Strategy:
The robot needs to solve for $I_{friction}$ dynamically.
-
The "Wiggle" Test: Periodically, the robot moves a limb in free air (zero load).
-
Zeroing: Any current measured during this "wiggle" is defined as pure friction/gravity noise.
-
Update: The system updates the bias term $\beta$ in the control loop.
This ensures that 0.1A of current always means "Touch," regardless of whether the robot is in a freezer or a desert.
6. What is the "reference touch" strategy and how does it self-correct?
To scale this economically, humanoids can use a Sparse Sensing Architecture. Instead of 1,000 sensors, use one high-precision load cell located at a "Calibration Station" (or even one on a single finger).
The Workflow:
-
The robot detects drift in its estimates.
-
It performs a "Reference Touch," pressing its finger against a known calibrated surface (or its own chassis).
-
It compares its Estimated Force (Current-based) vs. the Actual Force (Reference Sensor).
-
It calculates a correction factor ($Error_{\Delta}$) and applies it globally to all joints.
This allows the robot to self-heal its calibration drift without human intervention.
7. Why does actuator choice determine whether current sensing works?
Current-based force sensing only works if the actuator is mechanically predictable.
-
Rotary/Harmonic Drives: Often have complex, non-linear friction waves ("torque ripple") that make current sensing noisy and difficult.
-
Linear Actuators (ACME Screw): Devices like the FIRGELLI FA-BS16 use lead screws. The friction profile of a lead screw is highly linear and consistent compared to harmonic gears.
-
Low Backlash: Direct engagement means movement correlates instantly with motor rotation.
-
Consistent Current Draw: The 12V DC motor profile allows for clean current-to-force mapping.
-
Actuator suitability for current-based force estimation
| Actuator type | Friction profile | Backlash | Current-to-force mapping |
|---|---|---|---|
| Lead screw (e.g., FIRGELLI FA-BS16) | Linear, consistent | Low | Clean, predictable |
| Ball screw | Linear, low-friction | Very low | Clean, predictable |
| Harmonic / strain-wave drive | Non-linear with torque ripple | Very low | Noisy, requires filtering |
| Belt + pulley | Variable with tension/wear | Moderate | Drifts with belt condition |
8. How much force resolution does each humanoid task actually need?
A critical engineering insight is that different tasks require different resolutions of force sensing.
Sensing requirement by task level
| Task Level | Example tasks | Sensing needed | Hardware required |
|---|---|---|---|
| Level 1 — Collision detection | Did the arm hit something? | Coarse current monitoring | Motor current sensor (built into driver) |
| Level 2 — Calibrated force / tool use | Drilling, lifting, door pulls, gripping | Calibrated current sensing with adaptive bias | Predictable actuator (lead/ball screw), motor current sensor, optional reference load cell |
| Level 3 — Texture / slip detection | Distinguishing silk from sandpaper, micro-slip on a glass surface | Distributed tactile array | Electronic skin or capacitive/piezoresistive sensor patches |
-
Level 1 (Collision Detection): Did I hit something? (Requires coarse current monitoring).
-
Level 2 (Tool Use): Am I drilling with sufficient pressure? (Requires calibrated current sensing).
-
Level 3 (Texture/Slip): Is this silk or sandpaper? (Requires Tactile Skin).
95% of humanoid tasks (walking, lifting boxes, opening doors, using tools) fall into Levels 1 and 2. They do not require skin; they require Impedance Control.
Impedance Control Equation:
By adjusting the virtual Stiffness ($K$) and Damping ($B$) based on current feedback, the robot can act "soft" when handing an object to a human, or "stiff" when holding a heavy load, purely through software parameters.
9. Why is current sensing better than tactile skin for AI training and sim-to-real transfer?
For Artificial Intelligence (Reinforcement Learning), current sensing is often superior to tactile skin.
-
Data Structure: Tactile skin produces massive, high-dimensional, noisy point clouds that are hard to simulate.
-
Current Data: Current is a single scalar value ($I$). It is clean, low-latency, and numerically stable.
-
Sim-to-Real: It is much easier to simulate a motor's current draw in NVIDIA Isaac Sim or MuJoCo than it is to simulate the complex deformation of soft tactile skin. This leads to faster AI training times and more robust policy deployment.
What usually goes wrong with current-based force sensing?
- Thermal drift untracked: motor and gearbox warm up during use, grease viscosity changes, and the current baseline shifts. A grip calibrated in a 22°C lab can crush soft objects in a 5°C warehouse because the controller underestimates the friction the motor is fighting.
- Wrong actuator class: harmonic and strain-wave gearboxes produce torque ripple that looks like contact events in the current signal. Current-based force feedback only works cleanly on actuators with linear, predictable friction profiles — lead screws and ball screws.
- No periodic re-zeroing: friction terms drift slowly with wear, lubricant migration, and load history. Without a scheduled "wiggle test" in free air, the bias term β stays stale and force estimates degrade silently.
- Side loading on the actuator: a lead screw being side-loaded sees radically higher friction than one moving cleanly along its axis. The current-to-force model assumes axial loading. Side loads destroy the model — and the actuator — long before they bend anything.
- Treating Level 3 tasks as Level 2: current sensing cannot distinguish silk from sandpaper. Trying to extract texture from current draw produces unreliable behavior on any task that genuinely needs distributed tactile data.
How should you validate a current-based force-feedback system before trusting it?
- Run the wiggle test cold and hot. Move each limb in free air immediately after power-up (cold) and after 20–30 minutes of continuous motion (warm). Log the current baseline in both states. If the delta is not captured by your friction model I_friction(v, T), the model is incomplete.
- Verify against a known load. Press the end-effector into a calibrated load cell at three force levels (e.g., 5 N, 15 N, 50 N). Compare the current-derived force estimate to the measured force. Errors greater than ~10% across the working range mean k_sys or β needs re-tuning.
- Exercise the full velocity range. Friction is velocity-dependent. Repeat the calibration at slow, medium, and fast travel speeds — a model that only matches at one speed will fail in deployment.
- Compare against the simulation. Replay the same motion in Isaac Sim or MuJoCo with the modeled motor parameters. Current traces should match within reasonable bounds. Large divergences usually point to incorrect k_t or screw efficiency η values.
- Test failure response. Deliberately stall the actuator against a hard surface and confirm the controller detects the current spike and backs off within the intended response window — this is where impedance-control parameters K and B prove themselves.
Where current-based force sensing fits in real humanoid deployment
- Warehouse and logistics humanoids (lifting boxes, palletizing): force estimation prevents over-grip on deformable packaging and detects slip via current-rate-of-change.
- Domestic and service robots (opening doors, operating appliances): impedance control tuned by current feedback allows compliant interaction with mechanisms the robot has never seen before.
- Tool use (drilling, fastening, pressing): calibrated current sensing measures applied force at the tool tip without instrumenting the tool itself.
- Collaborative robots near humans: rapid current-spike detection enables stop-on-contact safety behavior without distributed skin sensors.
- AI training pipelines (Isaac Sim, MuJoCo): scalar current data simulates cleanly, making policies trained in simulation transfer to physical robots with far less domain randomization than tactile-skin policies require.
What does this mean for the future of humanoid robot design?
Humanoid robots will not reach mass adoption by becoming more complex hardware platforms; they will succeed by becoming smarter systems.
Replacing fragile, expensive tactile skin with robust, physics-based Current Sensing and Adaptive Calibration is not just a cost-cutting measure—it is an evolution in reliability. By utilizing high-quality linear actuators like the FIRGELLI FA-BS16 that offer predictable mechanical behaviors, engineers can build robots that feel the world through physics, not just sensors.
Ready to Engineer the Future?
Explore the FIRGELLI Micro Actuator Series—the linear actuator designed for the feedback-rich requirements of modern humanoid robotics
About the author: Robbie Dickson is the Founder and Chief Engineer of FIRGELLI Automations. Before founding FIRGELLI in 2002, he worked as an engineer at Rolls-Royce, BMW, Isuzu, and Ford. More on Wikipedia.
