NASA (robotics programs)
Last reviewed
Sources
19 citations
Review status
Source-backed
Revision
v4 · 3,304 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
19 citations
Review status
Source-backed
Revision
v4 · 3,304 words
Add missing citations, update stale details, or suggest a clearer explanation.
| NASA Robotics | |
|---|---|
| General information | |
| Full name | National Aeronautics and Space Administration |
| Founded | July 29, 1958 |
| Robotics center | Lyndon B. Johnson Space Center (JSC), Houston, Texas |
| Key lab | Dexterous Robotics Laboratory |
| Industry | Space exploration, Robotics |
| Notable robots | Robonaut R1, Robonaut 2, Valkyrie (R5) |
| Key AI systems | AEGIS autonomous targeting, Prithvi geospatial foundation model |
| Key partners | DARPA, General Motors, IBM Research, Apptronik (spinoff connection) |
| Website | nasa.gov/reference/jsc-robotics |
NASA (the National Aeronautics and Space Administration) is the U.S. space agency, founded in 1958, that builds robotic and artificial intelligence systems for space exploration, and its onboard autonomy software now lets Mars rovers select and analyze science targets without waiting for instructions from Earth. NASA's AEGIS system, which stands for Autonomous Exploration for Gathering Increased Science, has autonomously chosen rock targets on the Curiosity and Perseverance rovers and hit the most-desired material more than 93% of the time, compared with about 24% when targets are chosen blindly without onboard intelligence.[11][12] Alongside this mission autonomy, NASA's Johnson Space Center (JSC) in Houston, Texas, houses the Dexterous Robotics Laboratory, which has produced some of the most advanced humanoid robots ever built, including the Robonaut series and the Valkyrie (R5) full-size humanoid.[1][2]
NASA's robotics and AI programs have had a broad impact beyond space exploration. The agency's work on robotic manipulation, autonomous decision-making, and machine learning for science data has influenced the commercial robotics and AI industries. NASA co-developed the open-source Prithvi geospatial foundation model with IBM Research, is widely credited with the practical origin of the digital twin concept, and contributed the Kepler dataset used in a landmark deep-learning exoplanet discovery.[3][13][16][17] Most notably in robotics, Apptronik, the Austin-based company behind the Apollo humanoid robot, was co-founded by researchers who worked on the Valkyrie project at JSC, representing a direct technology transfer from NASA's humanoid robotics research into the commercial sector.[3]
NASA's most visible use of artificial intelligence in active missions is AEGIS (Autonomous Exploration for Gathering Increased Science), an onboard autonomous targeting system that lets a rover identify scientifically interesting rocks in its own camera images and aim a science instrument at them without a command from Earth. This matters because a one-way radio signal between Earth and Mars takes several minutes, so traditional target selection requires a full day of waiting for images to arrive, human analysis, and an uplinked command.
AEGIS was uploaded to the Curiosity rover in 2015 to drive its ChemCam Laser-Induced Breakdown Spectroscopy (LIBS) instrument, which vaporizes a tiny spot of rock with a laser and reads the resulting plasma to determine composition.[11] Over its first years of operation, AEGIS autonomously acquired data on more than 400 rock targets, and in a NASA technical assessment it selected the most desired target material greater than 93% of the time, versus roughly 24% for blind pointing without onboard intelligence.[11][12] An enhanced version was later added to the Perseverance rover for its SuperCam instrument: AEGIS-Lite was deployed to SuperCam on May 18, 2022, with an upgraded version arriving in February 2023.[12] As NASA describes it, AEGIS lets the instrument "autonomously identify and 'zap' rocks on Mars using its Laser-Induced Breakdown Spectroscopy (LIBS) technique."[12]
The practical benefit is increased science yield: AEGIS fills the rover's otherwise idle periods (for example, immediately after a long drive) with useful measurements, providing systematic baseline characterization of a site without round-trip communication to Earth.[11][12] Perseverance also runs additional autonomy software such as enhanced AutoNav for self-driving across terrain, part of a broader trend toward autonomous robotic operations on the Martian surface.[12]
Prithvi is an open-source geospatial foundation model co-developed by NASA and IBM Research, a temporal vision transformer pretrained on NASA Earth-observation satellite imagery so it can be fine-tuned for tasks such as flood mapping, wildfire burn-scar detection, crop classification, and land-use change. The first version, often called Prithvi-100M, has 100 million parameters and was pretrained on NASA's Harmonized Landsat and Sentinel-2 (HLS) dataset.[13][14]
Development of the model began in January 2023, and it was released publicly on August 3, 2023 via Hugging Face, where NASA described it as among the largest open-source geospatial AI foundation models at the time.[13][14] The collaboration extended beyond NASA and IBM to include Clark University's Center for Geospatial Analytics, the U.S. Geological Survey, and Oak Ridge National Laboratory.[13] Dr. Rahul Ramachandran, who manages NASA's IMPACT project, said "AI foundation models for Earth observations present enormous potential to address intricate scientific problems and expedite the broader deployment of AI across diverse applications."[13]
A larger successor, Prithvi-EO-2.0, was released on December 6, 2024 with 600 million parameters, roughly six times the size of the original. NASA and IBM reported it scored an average of 75.6% on the GEO-Bench benchmark, an 8 percentage-point improvement over the previous version.[15] The table below summarizes the Prithvi geospatial line.
| Version | Released | Parameters | Training data | Reported benchmark |
|---|---|---|---|---|
| Prithvi-EO-1.0 (100M) | August 3, 2023 | 100 million | Contiguous-U.S. HLS imagery | First-of-its-kind temporal vision transformer |
| Prithvi-EO-2.0 (300M / 600M) | December 6, 2024 | up to 600 million | Global HLS imagery | 75.6% avg. on GEO-Bench (8 pts over v1.0) |
NASA's Kepler space telescope produced the dataset behind one of the most cited applications of deep learning in astronomy. In 2017, Google AI engineer Christopher Shallue and University of Texas at Austin astronomer Andrew Vanderburg trained a neural network to spot the tiny dips in a star's brightness caused by a planet passing in front of it, using 15,000 previously-vetted signals from the Kepler catalogue to teach the network by example.[16][17]
The trained network identified true planets and false positives correctly about 96% of the time, and it recovered a previously-missed eighth planet, Kepler-90i, around the star Kepler-90, the first known star besides the Sun confirmed to host eight planets at the time.[16][17] NASA announced the result on December 14, 2017. "Machine learning really shines in situations where there is so much data that humans can't search it for themselves," Shallue said.[16] Vanderburg added, "It's like sifting through rocks to find jewels."[16] The work, published as "Identifying Exoplanets with Deep Learning," became an influential demonstration that convolutional neural networks could surface weak transit signals buried in large survey archives.[17]
The practical origin of the digital twin, a high-fidelity virtual model kept synchronized with a physical system, is widely traced to NASA's Apollo program in the 1960s, when the agency maintained ground-based simulators and "living models" that mirrored spacecraft in flight. The most famous example came during the 1970 Apollo 13 mission: after an oxygen tank exploded, engineers used physical and simulated models of the spacecraft on the ground to test survival procedures before relaying them to the crew.[18]
The term "digital twin" itself was popularized later. University of Michigan professor Michael Grieves presented the underlying model in the context of product lifecycle management around 2002 to 2003, and the specific phrase is often attributed to NASA's John Vickers.[18][19] Today NASA and aerospace partners use digital twins of launch vehicles, spacecraft, and ground systems for design, testing, and predictive maintenance, increasingly combining physics-based simulation with machine learning.[19]
NASA's involvement in robotics dates to the earliest days of space exploration. The agency developed telerobotic systems for satellite repair, planetary rovers for surface exploration, and robotic arms for spacecraft operations. Key early milestones included the development of manipulator arms for the Space Shuttle program and contributions to the Canadarm (Shuttle Remote Manipulator System, or SRMS), which was developed through a partnership between NASA and the Canadian National Research Council beginning in 1975.[4]
The Canadarm, a 15-meter robotic arm, flew on Space Shuttle missions from 1981 to 2011, deploying and retrieving satellites, supporting spacewalks, and later assisting with International Space Station (ISS) assembly. The Canadarm2 (Space Station Remote Manipulator System, or SSRMS), launched in 2001, is a 17-meter robotic arm that has been instrumental in ISS operations for over two decades. NASA's Johnson Space Center developed the flight software, training simulators, and operational procedures for both arms.[4]
Another significant Canadian contribution, the Special Purpose Dexterous Manipulator (SPDM), also known as Dextre, serves as a fine-manipulation robotic "hand" attached to Canadarm2 on the ISS. Standing 3.7 meters tall, Dextre performs delicate repair and maintenance tasks that would otherwise require astronaut spacewalks.[4]
The Robonaut project began in 1997 at NASA's Dexterous Robotics Laboratory at JSC, with the goal of developing a humanoid robot capable of assisting astronauts with a variety of manipulation tasks. The project was a collaboration between NASA and DARPA (the Defense Advanced Research Projects Agency). The core design philosophy was to create a robot with human-like dexterity that could use the same tools, handholds, and workstations designed for astronauts, eliminating the need for specialized robotic interfaces.[1][2]
Robonaut 1 (R1) was the first model, introduced in 2002 in two versions: R1A and R1B (a more portable variant). R1 featured a humanoid torso, two dexterous arms, and multi-fingered hands capable of grasping and manipulating tools. Several mobility platforms were developed for R1:
| Platform | Year | Description |
|---|---|---|
| Zero-G Leg | Early 2000s | Designed for station interior work in microgravity |
| Robotic Mobility Platform (RMP) | 2003 | Based on Segway technology for ground mobility |
| Centaur 1 | 2006 | Four-wheeled mobile platform for outdoor operations |
R1 was never flown to space but served as an essential testbed for the technologies that would enable its successor.[1]
In 2007, NASA signed an agreement with General Motors to jointly develop the next generation of humanoid robot technology. The collaboration produced Robonaut 2 (R2), which was publicly revealed in February 2010. R2 represented a major improvement over its predecessor.
| Specification | R1 | R2 |
|---|---|---|
| Speed | Baseline | 4x faster than R1 |
| Arm speed | N/A | Up to 2 m/s |
| Payload capacity | N/A | 18 kg (40 lb) |
| Grasping force | N/A | ~2.3 kg (~5 lb) per finger |
| Sensors | Limited | 350+ sensors |
| Processors | N/A | 38 PowerPC processors |
| Vision | N/A | 4 cameras (2 stereo + 2 auxiliary), infrared depth camera |
| Hand DOF | N/A | 12 DOF + 2 wrist DOF per hand |
| Control modes | Teleoperation | Autonomous + teleoperation |
On February 24, 2011, R2 launched to the International Space Station aboard Space Shuttle Discovery on the STS-133 mission, becoming the first humanoid robot sent to space. On August 22, 2011, R2 was powered up for the first time in orbit (a power system test with no movement). On October 13, 2011, R2 moved for the first time in space.[1][5]
In April 2014, a pair of climbing legs was delivered to the ISS via the SpaceX CRS-3 mission. Each leg contained seven joints and spanned 2.7 meters (9 feet) when fully extended. Instead of feet, the legs had "end effectors" consisting of a gripper and a camera, designed to grip handrails and allow R2 to move around the station's interior.[1][5]
However, R2 experienced persistent technical problems after the leg installation attempt. An intermittent power issue proved difficult to diagnose in orbit. In May 2018, R2 was returned to Earth aboard the SpaceX CRS-14 Dragon cargo spacecraft. Back at JSC, engineers discovered the root cause: a missing return wire in the power supply for the robot's computer chassis, which caused current to flow through an unintended "sneak circuit" that overheated a connector in R2's backpack.[1][5]
As of 2024, R2 is displayed at the Smithsonian's Steven F. Udvar-Hazy Center and has not been relaunched to the ISS.[1]
In late 2009, NASA proposed Project M, an ambitious plan to land a Robonaut 2 on the Moon within 1,000 days using a lander derived from the Lunar Atmosphere and Dust Environment Explorer (LADEE) spacecraft. The project was not approved for funding.[1]
Valkyrie (officially designated "R5" by NASA) is a full-size bipedal humanoid robot designed and built by the JSC Engineering Directorate. Building on the experience gained from the Robonaut program, the Valkyrie team designed and assembled the robot within a 15-month period beginning in October 2012.
| Specification | Detail |
|---|---|
| Height | ~1.87 m (~6 ft 2 in) |
| Weight | 129 kg |
| Degrees of freedom | 44 |
| Processors | Three Intel Core i7 Express CPUs |
| Battery life | ~1 hour |
| Hands | Each with a thumb and three fingers |
| Arms | 7 DOF each with actuated wrists |
| Legs | 6 DOF each |
| Neck | 3 DOF |
| Connectivity | Ethernet or WiFi |
Valkyrie was designed to compete in the 2013 DARPA Robotics Challenge (DRC) Trials, a competition that tested robots' abilities to perform disaster-response tasks. At the December 2013 DRC Trials, Valkyrie experienced network problems and failed to score any points. Despite this result, the robot demonstrated NASA's capability to rapidly design and build a sophisticated full-size humanoid.[6][7]
In mid-2015, NASA awarded two Valkyrie R5 robots to winning university teams:
| Recipient | Funding | Duration |
|---|---|---|
| Massachusetts Institute of Technology (MIT) | $500,000 | Two years |
| Northeastern University | $500,000 | Two years |
The university programs aimed to advance Valkyrie's autonomy and dexterous manipulation capabilities for potential future space exploration missions, where humanoid robots could perform tasks in advance of human arrival on other planets or serve as assistants to human crews.[7]
NASA's humanoid robotics research at JSC has had a direct impact on the commercial robotics industry through Apptronik, an Austin, Texas-based company that develops the Apollo humanoid robot.
Apptronik's CTO Nick Paine was researching robotic actuation for his Ph.D. at the University of Texas at Austin when he and his advisor Luis Sentis joined a team of engineers at Johnson Space Center to build Valkyrie for the DARPA Robotics Challenge. Paine had been influenced by several NASA papers on Robonaut 2's actuator design.[3]
When Paine and Sentis co-founded Apptronik in 2016 with CEO Jeff Cardenas, the company's first contract was a Small Business Innovation Research (SBIR) grant from NASA to develop liquid-cooled robotic actuator technology. Apptronik continued working with NASA through the Game Changing Development Program, and the Apollo humanoid robot was developed with direct NASA support as a continuation of the agency's work on humanoid robots for space applications.[3]
The Apollo robot is now deployed commercially at Mercedes-Benz factories and other industrial sites, representing one of the clearest examples of NASA technology transfer into commercial humanoid robotics.[3]
Beyond humanoid robots, NASA operates and has developed a wide range of robotic systems:
| System | Type | Mission |
|---|---|---|
| Canadarm / SRMS | Shuttle robotic arm | Payload deployment, satellite capture (1981 to 2011) |
| Canadarm2 / SSRMS | Station robotic arm | ISS assembly, maintenance, cargo handling (2001 to present) |
| Dextre / SPDM | Dexterous manipulator | Fine manipulation tasks on ISS exterior |
| Mars rovers | Planetary rovers | Sojourner, Spirit, Opportunity, Curiosity, Perseverance |
| Ingenuity | Mars helicopter | First powered flight on another planet |
| LEMUR | Climbing robot | Technology demonstrator for asteroid and cliff exploration |
NASA's robotics and AI programs have produced significant technology transfer to the broader robotics and machine-learning industries. The Robonaut program's work on dexterous manipulation, force-torque sensing, and impedance control has influenced the design of commercial collaborative robots. The Valkyrie program advanced the state of the art in full-body humanoid locomotion, whole-body control, and modular actuator design. On the software side, the AEGIS autonomous-targeting system demonstrated practical onboard autonomy for science operations, and the Prithvi collaboration showed how Earth-observation archives can train open foundation models for downstream geospatial tasks.[2][3][12][13]
The agency's approach to humanoid robotics, emphasizing human-compatible form factors that can use existing tools and workstations, has been adopted by multiple commercial humanoid robot companies. NASA's open publication of research results, open model releases, and its SBIR/STTR programs have facilitated the transfer of these technologies into the private sector, with Apptronik being the most prominent example of a commercial company directly descended from NASA's humanoid robotics research.[3]
NASA's robotics and AI programs continue to evolve as the agency prepares for future missions to the Moon under the Artemis program and eventual crewed missions to Mars. Humanoid robots are being considered for multiple roles in future exploration scenarios: they could be deployed to planetary surfaces before human arrival to set up habitats and infrastructure, serve as assistants during crewed missions to handle dangerous or tedious tasks, and maintain systems during the long transit periods between Earth and Mars when crew activity must be minimized.
Onboard autonomy will be central to those missions because communication delays make ground control impractical for time-sensitive decisions. The lessons learned from the Robonaut program's ISS deployment, particularly the challenges of operating humanoid robots in microgravity and the difficulties of diagnosing hardware failures remotely, have informed the design requirements for next-generation space robotics systems. The success of AEGIS in autonomously selecting science targets, and of foundation models like Prithvi in analyzing large Earth-observation datasets, points toward an increasingly AI-driven approach to both mission operations and science analysis.
NASA's continued investment in Small Business Innovation Research (SBIR) and Small Business Technology Transfer (STTR) programs ensures that the agency's robotics and AI expertise continues to flow into the private sector. Companies like Apptronik, which was directly seeded by NASA's humanoid robotics research, represent the agency's broader impact on the commercial robotics and AI ecosystem beyond its core space exploration mission.[3]