🤖 Robot Hardware

This blog summarizes key concepts from 6.4210 Fall 2023 Lecture 2: Let’s get you a robot! and supporting materials, focusing on robot arm hardware, the intricacies of simulation, and the evolving landscape of robot hands.

Course Link

1. Robot Arm Hardware: Evolution and Key Characteristics

🏭 Industrial Robots (Traditional)

  • Characteristics: Strong, fast, highly precise
  • Sensors: Primarily joint position sensors
  • Safety: Operate in isolated spaces; fault if trajectory deviates or if there’s an unexpected bump
  • Control: Mostly position-controlled
  • Limitation: Unsafe and impractical for home use

🧍‍♀️ Collaborative Robots (Cobots)

  • Vision: Designed for safe, close human interaction
  • Design Philosophy: “Huggability” — sense and react to contact
  • Common Arms:
    • Universal Robots (UR)
    • Kinova Jaco (electronics embedded at the base)
    • Kuka iiwa (robust under experimental stress)
    • Franka Emika Panda
    • ABB Cobots

📡 Sensor Progression for Huggability

  • Joint Position/Velocity: Basic
  • Force-Torque (Wrist): Detect contact at end-effector
  • Joint Torque Sensors: Detect torques along entire arm (iiwa, Panda)
  • Tactile Skin: Vision-based sensors or distributed force sensors — active research area

alt_text

⚙️ Why Most Robots Are Position-Controlled (Role of Transmissions)

Most robot arms today are primarily position-controlled, and while torque control is theoretically possible, practical challenges with hardware make position control the preferred approach. Position-controlled means the robot is told where to go (joint-wise), and it figures out how to apply the necessary forces to get there. It abstracts away the underlying physics so users don’t need to deal with low-level dynamics — which makes it easier to use, but less flexible than torque-controlled systems.

⚙️ A. Motor Characteristics and Desired Robot Motion

  • Electric motors are most efficient in a high-speed, low-torque regime.
  • Key relationships:
    • Current input ∝ Torque output
    • Voltage input ∝ Speed output
  • However, for manipulation tasks, robots need high torque and low speed — the opposite of motor characteristics.

⚙️ B. The Role of Transmissions

To bridge this mismatch:

  • Large transmissions (e.g., harmonic or planetary gearboxes) are used between the motor and robot joints.
  • These components are:
    • Complicated and lossy
    • Introduce backlash (gaps when reversing direction)
    • Suffer from flexing and bending under load
    • Add nonlinear friction and compliance

🛑 C. Why Torque Control Is Difficult Through Transmissions

  • Even if you control motor-side torque precisely (via current), it doesn’t translate reliably to joint-side torque.
  • Friction and compliance in the transmission distort the torque mapping.
  • This makes direct torque control at the output shaft inaccurate and difficult.

✅ D. The Solution: Joint-Side Position Sensing

To overcome these issues:

  • Place a position sensor on the robot’s output shaft (after the transmission).
  • Much easier and more reliable than using torque sensors everywhere.
  • A feedback loop is closed:
    • Controller sends motor current
    • Goal: Drive output shaft to desired position or velocity
  • PID controllers (Proportional-Integral-Derivative) are widely used:
    • Simple to implement
    • Works well in practice
🧩 Why Do We Need It?

Most robot arms use electric motors with gearboxes (e.g., harmonic drives, planetary gears) to convert high-speed, low-torque motor output into the high-torque, low-speed motion needed for manipulation.

However, gearboxes introduce issues like:

  • Backlash: A small “dead zone” when changing direction
  • Friction: Makes torque transmission unpredictable
  • Elasticity or compliance: Gear deformation under load

Because of these nonlinearities, measuring motor-side position or torque doesn’t give a reliable view of what the joint is actually doing.

🔍 Joint-Side Position Sensing Solves This

By mounting a sensor after the gearbox, directly on the joint, we measure the actual joint position regardless of what’s happening inside the transmission.

So the controller can:

  • Close a position feedback loop using the true joint angle
  • Drive the motor until the joint reaches the desired position, not just the motor
  • Ignore friction and flexibility of the gearbox
🔄 How It Works in a PID Loop

The joint encoder provides $( Q )$ (actual joint angle), and the controller compares it to the desired angle $( Q_d )$:

\[\tau(t) = K_P (Q_d - Q) + K_I \int (Q_d - Q) \, dt + K_D \frac{d(Q_d - Q)}{dt}\]

This generates a motor torque (or current) that moves the output shaft to the right position — not just the motor rotor.

🌀 E. Reflected Inertia and Why PID Works So Well

Surprisingly, fixed PID gains (set at the factory) work across many robot configurations. Why?

  • Reflected Inertia: Motor inertia, scaled by gear ratio squared, dominates joint dynamics.
    • Example: Gear ratio 160:1 → Reflected inertia = motor inertia × 160²
  • In robots like the Kuka iiwa, up to 85% of a link’s inertia may come from moving the motor, not the link.
  • Implication:
    • Much of the robot’s work is moving its own motor mass
    • This stabilizes dynamics across different tasks and configurations
    • Fixed PID gains remain effective, avoiding complex gain scheduling

🔁 Alternative: Direct Torque Control Designs

Robots focused on torque control include:

  • Direct Drive Robots: Low gear ratios (or no gearboxes), large motors
  • Series Elastic Actuators (SEA): Spring in series with motor to measure torque
    • Example: Baxter

These represent a different design philosophy, often used for compliant or dynamic interaction but are less common than traditional position-controlled arms.

PID

PID (Proportional-Integral-Derivative) control is one of the most widely used feedback control mechanisms in robotics, particularly in position-controlled robot arms. It’s known to work “shockingly well” — often allowing engineers to set fixed control gains once in the factory, despite variations in robot configuration and payload.

PID control works by:

  • Measuring the current system output (e.g., joint angle Q)
  • Comparing it to the desired value (Q_desired)
  • Computing a control command (e.g., motor current/torque) to reduce the error over time

This forms a feedback loop where sensors track the actual state, and the controller continually adjusts commands to reach and maintain the target.

A. Proportional (P) Term

  • Formula: Kp * (Q_desired - Q)
  • Meaning: Directly proportional to current error
  • Behavior:
    • Larger error → stronger response
    • Kp (proportional gain) determines reaction strength
  • Analogy: Like a spring pulling toward the goal

B. Integral (I) Term

  • Formula: Ki * ∫(Q_desired - Q) dt
  • Meaning: Accumulates past errors over time
  • Benefits:
    • Eliminates steady-state error
    • Helpful when proportional term alone can’t reach the exact setpoint (e.g., balancing against gravity)
  • Ki (integral gain) scales this effect

C. Derivative (D) Term

  • Formula: Kd * d(Q_desired - Q)/dt
  • Meaning: Reacts to rate of error change
  • Benefits:
    • Damps oscillations
    • Anticipates overshoot and smooths motion
  • Kd (derivative gain) sets the sensitivity to change

🔧 Application in Robot Arms

  • Used to command motor current to drive the output shaft of a joint to the desired position
  • Q is typically a relative joint angle
  • Can be augmented with feedforward terms, such as:
    • Gravity compensation (e.g., computing torque required to counteract gravity)
    • Trajectory feedforward (for faster or smoother motion)

💡 Why PID Works So Well (Even with Fixed Gains)

One surprising fact: PID gains often don’t need to change across robot configurations. Why?

The Reason: Reflected Inertia

  • Electric motors are paired with high gear ratio transmissions (e.g., 160:1 in Kuka iiwa)
  • The reflected inertia of the motor is scaled by the square of the gear ratio
  • This reflected inertia:
    • Can dominate the robot’s joint dynamics
    • Makes the joint “feel” very similar regardless of load or configuration
  • Result:
    • PID control becomes less sensitive to external changes
    • Fixed gains like Kp, Ki, and Kd remain effective across a wide range of tasks

🖥️ Implementation Details

  • Modern PID controllers are digitally implemented, often running at high control rates (e.g., 10 kHz)
  • Advantages:
    • High-frequency control loop ensures responsive and stable behavior
    • Digital configuration allows easy tuning and monitoring
  • In early days, analog circuits were used, but are now obsolete in most applications

✅ Summary

Component Role Common Gain
Proportional Reacts to current error Kp
Integral Accumulates past error Ki
Derivative Reacts to rate of error change Kd

PID control offers a simple yet powerful strategy to control robotic actuators. With reflected inertia helping to stabilize the system’s behavior, fixed PID gains can perform robustly across varied tasks — a key reason why this approach remains standard in modern robot arms.

🌀 Reflected Inertia

  • Motor inertia scaled by gear ratio squared
  • Can be dominant in robot joint dynamics (e.g., ~85% on Kuka iiwa link)
  • Allows fixed PID gains to work well across configurations

🧠 Direct Drive Robots

  • Approach: Eliminate large gear ratios (gear ratio ≤ 10)
  • Motors: Large direct drive motors or OutRunner motors
  • Benefit: Enable true torque control via motor current

🔧 Achieving Joint Torque Sensing

  • Strain Gauges: High stiffness sensors (e.g., Kuka iiwa)
  • Series Elastic Actuators (SEA):
    • Spring in series with motor
    • Example: Baxter
    • Lower bandwidth than strain gauges
  • Hydraulic Actuators: Atlas v1; use differential pressure
  • Impact: Enables impedance control, gravity compensation, and safe human-robot interaction

2. 🤖 Drake’s MultibodyPlant: The Core of Robot Physics Simulation

MultibodyPlant is Drake’s physics engine, serving as the core component responsible for simulating the dynamics of robotic systems.

🔧 Core Function

  • Takes torque commands as input.
  • Computes:
    • Body angles
    • Joint angles
    • Positions
    • Velocities
    • Accelerations
  • Models the robot’s rigid-body physics, capturing how it behaves under applied forces without any high-level control logic.

🧩 Interaction with Other Components

A. SceneGraph

  • Drake’s geometry engine that works alongside MultibodyPlant.
  • Responsible for:
    • Rendering robot visuals
    • Collision detection
    • Mesh distance queries

B. Robot Description Files

  • MultibodyPlant is populated using robot description files such as:
    • SDF (Simulation Description Format)
    • URDF
  • These files define:
    • Links and joints
    • Inertial properties
    • Geometry meshes
    • Material characteristics

⚠️ Raw Simulation vs. Real Behavior

  • A standalone MultibodyPlant with zero torque input will simulate the robot falling under gravity.
  • But real robots (e.g., Kuka iiwa) hold their position when powered on, even without external commands.

Real robots include low-level controllers (e.g., gravity compensation, braking), not modeled by physics alone.

🧠 Simulating Low-Level Control

To replicate real behavior:

  • Add controllers like InverseDynamicsController to the simulation.
  • These:
    • Read the estimated current state
    • Compare it to a desired state
    • Output torques to hold position or track motion
  • Enables simulation of:
    • Gravity compensation
    • Posture maintenance
    • Joint-level control

⚙️ Reflected Inertia Support

MultibodyPlant supports:

  • Rotor inertia
  • Reflected inertia

Reflected Inertia Explained:

  • In robots with large gear ratios, motor inertia (small in absolute terms) becomes amplified by the square of the gear ratio.
  • This has a major impact on joint dynamics, especially for robots like the Kuka iiwa with high gear ratios (e.g., 160:1).

Limitation:

  • Many robot description formats don’t include fields for reflected inertia.
  • This can result in inaccurate simulations if not accounted for.

✅ Summary

🧩 Core Components of a Simulator (e.g., Drake)

  • MultibodyPlant (Physics Engine): Computes dynamics from torque input
  • Scene Graph (Geometry Engine): Collision detection, rendering, distance queries
  • Robot Description Files (e.g., SDF):
    • Text files with links, meshes, inertia, materials
    • Prone to bugs → need visualization

⚙️ Low-Level Controller Modeling

  • Real robots use control cabinets (e.g., gravity compensation)
  • Simulators must include mathematical models of these controllers
    • Often implemented as inverse dynamics
    • Needed for realistic and accurate behavior

🚧 Reflected Inertia in Simulation

  • Drake supports rotor/reflected inertia
  • Most common formats (URDF/SDF) lack fields for this
  • Omission leads to less realistic simulation
Feature Description
MultibodyPlant Simulates rigid-body dynamics of robots
Inputs Torque commands
Outputs Joint/body states (position, velocity, acceleration)
Paired With SceneGraph for geometry and collision
Needs Controllers To model realistic robot behavior like holding position
Supports Rotor and reflected inertia (when specified)

MultibodyPlant is essential for high-fidelity robotics simulation, but its effectiveness depends on combining it with low-level controllers and accurate modeling of transmission effects like reflected inertia.

3. Robot Hands: Dexterity vs. Utility

✋ Dexterous Hands

  • Examples: Shadow, Allegro, Sandia
  • Pros: Capable of complex manipulation (Rubik’s cube, in-hand regrasping)
  • Cons: Fragile, hard to calibrate, research-grade only

✂️ Simple Grippers

  • Philosophy: “Big brain, simple claw”
  • Examples: PR2 with 2-finger gripper, Schunk WSG
  • Utility: Perform most household tasks robustly
  • Feature: Force control, interchangeable fingers, tactile integration

🧪 Innovative Hand Designs

  • Granular Jamming Grippers: Vacuum-packed coffee grounds
  • Soft Hands: Pneumatic actuators + vision-based tactile sensing
  • Underactuated Hands: Cables and springs compensate for fewer actuators (e.g., iHy hand)

🔮 Future Trends

  • Robust Dexterous Hands: Many humanoid startups are tackling the challenge
  • High-Speed Hands: Ishikawa hand performs dynamic tasks like dribbling and regrasping

4. Mobile Manipulators

🧍‍♂️🦾 Mobile + Arm Systems

  • Examples:
    • PR2, Fetch, HSR
    • Google Everyday Robot
    • Boston Dynamics Spot
    • TRI’s custom platforms

❗Challenges

  • Hard to commercially obtain a PR2-class mobile manipulator
  • Many research groups build custom hardware

🐶 Spot as a Mobile Base

  • Has arm, but limited field of view (must use hand camera to see workspace)