Robonaut 2

NASA/Keegan Barber. Public domain (NASA / US government work).

Overview

Robonaut 2 is a two-armed anthropomorphic torso developed jointly by NASA Johnson Space Center and General Motors [4]. It is the second of the Robonaut line; Robonaut 1 was a 43 degree-of-freedom design with 7 joints per arm, a 3 degree-of-freedom waist and 14 per hand-and-wrist [4]. NASA’s own project page gives each R2 arm 7 degrees of freedom and a static hold capacity of 20 lb [12].

Specifications

Parameter	Value	Source
Independent degrees of freedom	42	[1]
Motors	50	[1]
Sensors	more than 350	[1]
Mass	about 300 kg with climbing legs
Arm joints	5 per arm, series elastic
Wrist	2 DOF per wrist
Hand	12 DOF per hand
Head	3 DOF
Waist	1 joint, series elastic
Climbing legs	7 DOF per leg, 14 total, series elastic

Rows with no marker are from [2].

Mission profile

Parameter	Value
Host	International Space Station, fixed base stanchion
Launch	STS-133, February 2011
Activated on orbit	August 2011
Task board	modular, with powered, unpowered, EVA, stowage and soft goods panels
Station tasks demonstrated	handrail cleaning, air velocity measurement

Source: [3]. R2 arrived on the station on the same STS-133 flight and was mounted as a fixed-base upper body on a stanchion in the US Lab, with climbing legs planned to follow once the task board program was under way [7]; the first crew teleoperation of the arms came in April 2013 [3]. Physical dimensions reported outside the NASA design papers put the torso at 101.6 cm from waist to head and 78.74 cm across the shoulders, at 149.7 kg without the climbing legs, able to lift 9.07 kg in Earth gravity [8].

Actuation

Joints use brushless DC motors through harmonic drives with series elastic actuation, each joint carrying a planar torsion spring sized individually for that joint and a 19 bit absolute position sensor across it [1]. Torque is measured as spring deflection rather than by a separate load cell, so the same element that provides compliance provides the torque signal.

Robonaut 1 established the torque range the design has to cover: a modular joint family in three torque scales spanning 13.6 to 271 N m (10 to 200 ft-lb), with better than 20 Hz bandwidth in the strain-gauge torque loop [4]. The hand requirement came from the EVA interface standard, 89 N (20 lbf) of force and 3.4 N m (30 in-lb) of torque to remove and install an orbital replacement unit [11]. R1 packaged all fourteen hand and wrist motors, the motor control and power electronics and the hand wiring into a forearm 102 mm (4 in) in diameter at its base and about 203 mm (8 in) long [4].

A closed-loop hand-eye calibration procedure applied to the R1 unit at JSC, fitting Denavit-Hartenberg parameters to stereo camera observations of a tracked sphere, cut the mean mismatch between the kinematic model and the vision system from 13.75 cm to 1.85 cm on the calibration data and to as little as 2.35 cm on held-out configurations [9]. The result registers the two onboard models to each other, not to an external ground truth, and rests on sample sizes of a few dozen to a few hundred configurations [9].

Hand

Each hand has five fingers: index and middle at 3 DOF, a 4 DOF thumb, and ring and little fingers at 1 DOF each, 12 DOF in total plus a 2 DOF wrist [1]. Fourteen six-axis load cells are distributed through the fingers. The R2 hand reproduces about 90 percent of the Cutkosky grasp taxonomy against about 50 percent for R1 [1], [5], the improvement coming largely from the four-jointed thumb, which is made stronger than the fingers it opposes [1].

The hand is designed against the EVA tool set rather than against purpose-built robot interfaces, and has been operated with sockets, wrenches, drills, wire strippers and flashlights on the ground [1] and with handrails, cables and soft goods on the station.

Tactile sensing is carried in a glove rather than in the hand structure. The first generation used 19 Interlink force-sensing resistors of 0.25 inch diameter; the second used 33 pads of quantum tunnelling composite, whose resistance falls from about 10 MOhm to about 1 Ohm across a 0.1 to 10 N force range [5]. With that glove the robot grasped a drill anywhere in a 2 by 1 foot workspace and within plus or minus 5 degrees about two axes.

Compute and data

Element	Specification
Central processors	2 PowerPC cards in a Compact PCI chassis
Operating system	VxWorks
Software framework	ControlShell
Joint controller	Superdriver, FPGA plus PowerPC, per node
Distributed nodes	5 in the upper arm, 2 in forearm and hand
Serial bus	MLVDS at 50 Mbit/s
Main arm cable	16 conductors, against more than 100 on R1

Source: [1].

Commutation is configurable between six-step and space vector [1]. Reducing the arm cable from over 100 conductors to 16 is a direct consequence of moving commutation and current control into the joint: the cable carries power and a serial bus rather than one pair per phase per motor.

Control

Control loops are nested by rate. Current and series-elastic torque control run at 10 kHz in the joint controller [1]; the embedded impedance loop runs at 5 kHz across four consecutive control loops, with 50 Hz the minimum acceptable rate in a degraded mode [2]; wrist and finger control runs near 500 Hz on the brainstem processor.

The control law is a dual-priority hierarchical impedance scheme: operational-space impedance governs the primary linear and angular task, and joint-space impedance projected into the null space governs the remaining freedom [2]. Gravity, Coriolis and inertia compensation are computed with a recursive Newton-Euler algorithm, and inverse kinematics uses an iterative multi-node pseudo-inverse. Trajectories are trapezoidal in velocity and accept nonzero initial and final conditions.

Impedance is retuned per task rather than fixed. Climbing tests used a stiff setting of natural frequency 20 and damping ratio 1.2 with a 100 torque-unit limit on the base leg and 20 on the reaching leg, against a soft setting of natural frequency 0.5 and damping ratio 1.5 for closing a gripper [2].

A 14 degree-of-freedom hand does not map onto a joystick or a hand-held haptic controller, which motivated a ground demonstration of markerless finger tracking as a teleoperation input: the operator’s hand posture drove the simulated R2 hand directly, with color cues substituting for the missing haptic and contact feedback and the robot advancing only on an explicit operator commit rather than continuous mirroring, to absorb the light-time delay to the station [10]. The demonstration ran entirely against a simulator with idealized grasping and reported no measured task times or error rates, so it establishes the interface concept and not a performance figure [10].

Safety

R2 is certified to work in the same volume as crew without a physical emergency stop, which is replaced by a motion-stop button [1]. Basic torque control limits the force the robot can apply, and separate software monitoring loops read the joint torque sensors and the arm force sensors independently, cutting motor power when either exceeds its limit; further routines on several processors check the health of both loops, giving a triple-redundant force limiting system. The station certification rests on a fail-safe fault containment region approach, in which two demonstrably independent means of failing safe cover each hazard: for inadvertent contact these are the shoulder force sensors and the per-joint absolute position sensors that measure torque in each joint [3]. The safety functions run entirely on the brainstem processors inside the torso, so they remain active if the link to an outside computer is lost.

Mobility

Climbing legs add 7 series elastic degrees of freedom per leg [2]. In gravity-offload testing a leg reached more than 1 m with better than 1 cm positioning error under full offload. The station program ran the upgrade in phases: intravehicular mobility first, with the climbing legs and a battery backpack, and extravehicular capability after that [3].

Operations on the station

R2 worked a purpose-built task board that flew with it, of modular design with interchangeable panels [3]. The powered panel carries push buttons, two and three way switches, a guarded rocker and a guarded power switch; the unpowered panel carries valves and connectors of the kinds found on station hardware; an EVA panel carries a handrail, a tether point and common EVA fittings; a stowage panel carries restraints; and a soft goods panel carries a fabric-covered box. The board was used repeatedly to practice and refine the motion scripts for generic tasks such as flipping switches, cleaning handrails and handling soft goods, and to test the machine vision that lets R2 verify whether a switch actually moved. Demonstrated station tasks include handrail cleaning and air velocity measurement with the Velocicalc air flow meter, both routine crew chores. Head stereo cameras on a 3 DOF pan and tilt provide the imagery [4], with low-rate visual servoing used to acquire handrails [2].

Extravehicular limits

R1 motors were thermal-vacuum rated and the design targeted an 8 hour EVA thermal endurance [4]. Vacuum operation would also expose the harmonic drives and bearings to the lubricant constraints that govern all orbital mechanisms, where vapor pressure rather than viscosity sets the fluid choice [6]. R2 never operated outside the pressurized volume; it was returned to Earth for repair and did not fly again.

What is not established

The hand-eye calibration figures for the R1 hand register the robot’s own kinematic and vision models against each other rather than against an external ground truth, on sample sizes of a few dozen to a few hundred configurations [9]. The finger-tracking teleoperation concept for the R2 hand was demonstrated only in simulation, with no measured task times, error rates or characterization of the tracker’s own accuracy, and it was never connected to flight hardware [10]. No EVA thermal or vacuum test of R2 itself is reported; the thermal-vacuum rating and 8 hour endurance figure belong to the R1 design [4], and R2’s own vacuum exposure was never demonstrated before it was returned to Earth.