DARPA Robotics Challenge
Last reviewed
Sources
38 citations
Review status
Source-backed
Revision
v3 ยท 6,792 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
38 citations
Review status
Source-backed
Revision
v3 ยท 6,792 words
Add missing citations, update stale details, or suggest a clearer explanation.
The DARPA Robotics Challenge (commonly abbreviated DRC) was a prize competition organised by the United States Defense Advanced Research Projects Agency (DARPA) that ran from October 2012 through June 2015, awarding a 2 million US dollar grand prize to the team whose robot best completed eight disaster-response tasks.[6][7] Team KAIST of South Korea won the June 2015 Finals with the DRC-Hubo robot, completing all eight tasks in 44 minutes and 28 seconds, ahead of two perfect-scoring rivals that ran slower.[23] The competition asked international teams to develop semi-autonomous ground robots, mostly humanoids, capable of performing complex tasks in the kind of degraded, human-engineered environments that arise after natural and industrial disasters.[6] It was directly motivated by the March 2011 Fukushima Daiichi nuclear accident, where the inability of any existing robot to enter the reactor buildings and operate manual controls became a global wake-up call for the robotics community.[30]
The DRC unfolded across three competitive events: a cloud-based Virtual Robotics Challenge (VRC) in June 2013, the DRC Trials at the Homestead-Miami Speedway in December 2013, and the DRC Finals at the Fairplex in Pomona, California in June 2015.[6] The Finals awarded a 2 million US dollar grand prize to the winning team, with 1 million for second and 500,000 for third, and an overall prize purse of 3.5 million dollars.[7] Team KAIST of South Korea won with the DRC-Hubo robot, completing all eight tasks in 44 minutes and 28 seconds.[23] The Florida Institute for Human and Machine Cognition (IHMC) finished second with a modified Atlas named Running Man, and Carnegie Mellon's Tartan Rescue placed third with the four-limbed CHIMP robot.[7]
The DRC is widely regarded as the most influential humanoid robotics competition of the early 21st century. It produced the open-source Gazebo simulator that still powers much of the field, established the original electric Atlas as the canonical research humanoid, accelerated whole-body locomotion and manipulation research, and seeded the engineering teams behind many of the humanoid robot companies that emerged in the 2020s.[37] It is also remembered for hours of video footage of large humanoids falling spectacularly during the Finals, a body of failure data that drove the next decade of improvements in balance control, fall recovery, and reliability engineering.[21]
| Field | Detail |
|---|---|
| Competition name | DARPA Robotics Challenge (DRC) |
| Organiser | Defense Advanced Research Projects Agency (DARPA) |
| Program manager | Dr. Gill Pratt |
| Announced | April 2012 |
| Program duration | October 2012 to June 2015 |
| Major events | Virtual Robotics Challenge (June 2013), DRC Trials (December 2013), DRC Finals (June 2015) |
| Trials venue | Homestead-Miami Speedway, Homestead, Florida |
| Finals venue | Fairplex, Pomona, California |
| Grand prize | 2,000,000 USD |
| Second prize | 1,000,000 USD |
| Third prize | 500,000 USD |
| Total purse | 3.5 million USD |
| Winning team | Team KAIST (South Korea) |
| Winning robot | DRC-Hubo |
| Winning time | 44 minutes 28 seconds, 8 of 8 tasks |
| Standard reference platform | Atlas, built by Boston Dynamics |
| Standard reference simulator | Gazebo, by the Open Source Robotics Foundation |
| Scientific motivation | 2011 Fukushima Daiichi nuclear disaster |
The DRC was conceived in the months following the March 2011 Tohoku earthquake and the subsequent meltdowns at the Fukushima Daiichi nuclear power plant.[30] In the first hours of the accident, plant operators were unable to vent hydrogen from the reactor buildings because the controls required manual intervention inside high-radiation zones. Workers attempted to reach the valves on foot but were turned back by lethal dose rates. Existing emergency response robots, including Japanese, American, and European designs that had been developed at considerable expense over the previous two decades, were unable to climb the stairs, open the doors, or operate the simple manual valves inside the plant.[30] The hydrogen explosions that followed contaminated a wide area of northeast Japan and underscored a sobering global gap between the public image of advanced robotics and what robots could actually do in a real disaster.[30]
DARPA responded by chartering a new program. Dr. Gill Pratt, a former MIT professor and a long-time figure in legged-robot research, was named program manager. In a series of public talks and interviews in late 2011 and early 2012, Pratt argued that disaster response was the right forcing function for the next leap in robot capability because it required the simultaneous integration of perception, manipulation, locomotion, and supervised autonomy under stressful, unpredictable conditions. The official DARPA framing of the challenge, repeated in nearly every program document, was the development of robots capable of executing complex tasks in dangerous, degraded, human-engineered environments.[6]
The program was first publicly announced on April 10, 2012, with the formal kickoff meeting held in October 2012.[24] Pratt told IEEE Spectrum at the time that his target performance bar for the Trials was the equivalent of a one-year-old child operating in a real environment.[24] He framed that goal as deliberately humble. The point of the program was to push the field toward reliable, real-world humanoid behaviour, not to stage a pre-baked demonstration. Reflecting on the progress between the two events, Pratt later compared watching the 2012 program kickoff to "watching grass grow" and the 2015 Finals to spectating a golf match, a measure of how far the robots had come.[22] To remove any temptation toward open-loop choreography, the Finals would impose tightly controlled communication degradation, force robots to perform a surprise task announced only on the day of competition, and impose a single-attempt format for each run.[38]
DARPA divided participants into four tracks designed to accommodate different funding profiles and engineering strategies. Tracks A and B were funded by DARPA. Tracks C and D were not.[24]
DARPA contracted Boston Dynamics to build seven hydraulic Atlas humanoids for delivery to the top VRC finishers, six of which were awarded to teams and one held in reserve.[5] The contract was reported at approximately 10.9 million dollars. Atlas became the de facto reference humanoid of the program, used by IHMC, MIT, Worcester Polytechnic Institute, TRACLabs, TROOPER, VIGIR, and several other Track B teams.[5] The Open Source Robotics Foundation (OSRF), led by Brian Gerkey and Nate Koenig, was contracted to extend the Gazebo simulator into the DRC Simulator (drcsim), adding higher-fidelity contact physics, sensor simulation, and cloud deployment so that VRC competitors could submit controllers as Linux containers.[27]
The deliberate split between funded hardware tracks and a software-only competition was an important policy choice. By absorbing the cost of Atlas hardware for some teams and providing a free, accurate simulator for the rest, DARPA lowered the barrier to entry for software-focused researchers and made it possible for university labs without hardware budgets to compete on a level field. The drcsim package and its associated Atlas model remain widely used in robotics teaching and research more than a decade later.[27]
The Virtual Robotics Challenge ran from June 17 to 21, 2013. Twenty-six teams submitted controllers that were executed against a simulated Atlas inside drcsim, hosted on Amazon Web Services.[16] The simulated robot had to complete three tasks: drive a utility vehicle through an obstacle course, walk over rough terrain with rubble, and connect a fire hose to a standpipe and turn a valve.[16] Communications between the simulated robot and the team operator were degraded to mimic real disaster conditions, with bandwidth and latency penalties applied dynamically.[16]
IHMC took first place in the VRC, followed by Worcester Polytechnic Institute, MIT, Team TRACLabs, JPL, TORC Robotics, Case Western Reserve University, ViGIR, K-Team, and the Trooper team led by Lockheed Martin. The top six teams were awarded a physical Atlas humanoid each, with one extra unit reserved by DARPA. Several VRC teams later combined or transferred their Atlas allocations, which is why the eventual list of Atlas-using Track B teams at the Trials and Finals does not perfectly match the original VRC ranking. The cloud-only format also opened the door for teams from countries that lacked the heavy industrial infrastructure needed to build hydraulic humanoids, broadening the geographic spread of Finals participation.
The DRC Trials were held on December 20 and 21, 2013 at the Homestead-Miami Speedway, the same NASCAR oval used for the Ford EcoBoost 300.[3] DARPA partitioned the infield into eight separate task zones, each with its own grandstand and overhead scoreboard.[4] Sixteen teams qualified to compete with physical hardware, drawn from the Track A hardware groups, the VRC winners (now using Boston Dynamics Atlas), and a handful of Track D entries.[3]
Each team had 30 minutes per task, with one attempt per task spread across the two days.[4] Communications between operator and robot were intentionally throttled to 9600 baud through a low-bandwidth link with periodic blackouts, while a higher-rate sensor link operated only during predefined windows.[4] Scoring awarded one point per completed task, plus bonus points for completing the task without using a safety harness or operator intervention. The maximum possible score was 32.[3]
SCHAFT, the Tokyo-based humanoid startup founded by Yuto Nakanishi and Junichi Urata, dominated the Trials. The custom SCHAFT robot, built around proprietary high-torque liquid-cooled actuators, scored 27 points.[25] IHMC came in second with 20 points using their loaner Atlas, and Tartan Rescue (CMU NREC) finished third with 18 points using the four-limbed CHIMP robot.[3] MIT placed fourth with Atlas at 16 points, and NASA JPL's RoboSimian placed fifth with 14 points. Several teams scored zero. The eight Trials tasks were:
| # | Task | Description |
|---|---|---|
| 1 | Vehicle | Drive a Polaris Ranger XP 900 around a course of cones, then exit the driver's seat. |
| 2 | Terrain | Walk across rough cinder-block terrain that progressed from flat to angled to chevron-stacked. |
| 3 | Ladder | Climb a 60-degree industrial ladder with railings. |
| 4 | Debris | Move a series of wooden 2x4 beams blocking a doorway. |
| 5 | Door | Open a sequence of three different doors: a push door, a pull door, and a weighted spring door. |
| 6 | Wall | Use a power tool to cut a triangular hole through a drywall panel. |
| 7 | Valve | Locate and rotate three different industrial valves in a designated direction. |
| 8 | Hose | Pick up a fire hose, walk it to a standpipe, and screw the connector on. |
Four days after the Trials, on Christmas Eve 2013, Google announced that it had acquired SCHAFT through its Replicant robotics group, joining a wave of acquisitions that also included Boston Dynamics, Industrial Perception, Redwood Robotics, Bot & Dolly, Meka Robotics, and Holomni.[25] SCHAFT subsequently withdrew from the funded portion of the DRC, with founder Yuto Nakanishi explaining that the team wished to develop commercial products free of military funding constraints.[25] SCHAFT continued to operate semi-independently inside Google for several years, demonstrated a new bipedal robot at the New Economic Summit conference in Tokyo in 2016, and was finally shut down by Alphabet in late 2018 after attempts to find a buyer fell through.[28]
The DRC Finals were held on June 5 and 6, 2015 at the Fairplex exposition grounds in Pomona, California.[2] Twenty-three teams brought robots to compete, drawn from a pool of 25 invited entries.[2] The audience over the two days was estimated at more than 17,000 people, with the event branded as a robotics festival rather than a strictly closed competition.[33] NASA and DARPA staff organised the surrounding Robotics Expo with university research demos, autonomous car displays, and a separate humanoid showcase.[2]
DARPA reset the rules to push robots harder than at the Trials. Each team got two complete one-hour runs, one on each day, instead of separate per-task attempts.[7] All eight tasks had to be completed back-to-back inside that single hour, with the clock running continuously. A 10-minute reset penalty applied if a team requested human intervention; falls counted as resets.[7] Communications were degraded much more aggressively. While the robot was outside the building, teams had a 300 Mbit/s telemetry link from robot to operator. Once the robot crossed an indoor threshold (after climbing the stairs), the link was rate-limited to the same 300 Mbit/s but interspersed with full blackouts that lasted up to 30 seconds at a time.[38] A separate 9600 baud command link from operator to robot remained on at all times, providing roughly 500 ms ping latency, deliberately too slow for traditional joystick teleoperation.[38]
The Finals course consisted of eight tasks performed in the following order:[7]
| # | Task | Description |
|---|---|---|
| 1 | Drive | Drive a Polaris Ranger XP 900 utility vehicle along a curved course past cones to a designated stop point. |
| 2 | Egress | Get out of the driver's seat of the Polaris and stand on the ground unaided. |
| 3 | Door | Open a 33-inch industrial door with a lever handle and walk through it into the building. |
| 4 | Valve | Locate a circular valve handle 10 to 40 cm in diameter and rotate it 360 degrees counterclockwise. |
| 5 | Wall | Pick up one of two cordless drills from a shelf and use it to cut a triangular hole in a drywall panel along a marked outline. |
| 6 | Surprise | Perform a manipulation task announced on the morning of the run. On Day 1, the surprise was unplugging an electrical plug from one socket and plugging it into another. On Day 2, the surprise was pulling a lever to actuate a switch. |
| 7 | Rubble | Either traverse a debris field of cinder blocks placed at irregular angles or, alternately, push aside scattered debris and walk through the cleared lane. |
| 8 | Stairs | Climb a flight of four industrial steps to reach a final platform. |
The stairs replaced the Trials ladder, which had proven too difficult for most platforms. The drive task no longer included an explicit obstacle slalom, but the egress task was singled out by many teams as the hardest single element on the course because of the awkward sit-to-stand transition required from inside the open driver's cab.
Twenty-three teams ran the course. The top ten finishers, with scores out of 8 points and the elapsed time of their best run, were:[7]
| Rank | Team | Country | Robot | Points | Best Time |
|---|---|---|---|---|---|
| 1 | Team KAIST | South Korea | DRC-Hubo | 8 | 44:28 |
| 2 | IHMC Robotics | United States | Running Man (Atlas) | 8 | 50:26 |
| 3 | Tartan Rescue (CMU NREC) | United States | CHIMP | 8 | 55:15 |
| 4 | NimbRo Rescue | Germany | Momaro | 7 | 34:00 |
| 5 | Robosimian (NASA JPL) | United States | RoboSimian | 7 | 47:59 |
| 6 | MIT | United States | Helios (Atlas) | 7 | 50:25 |
| 7 | WPI-CMU | United States | WARNER (Atlas) | 7 | 56:06 |
| 8 | DRC-Hubo at UNLV | United States | DRC-Hubo | 6 | 57:41 |
| 9 | TRACLabs | United States | Hercules (Atlas) | 5 | 49:00 |
| 10 | AIST-NEDO | Japan | HRP-2Kai | 5 | 52:30 |
Team KAIST's victory hinged on a design choice that distinguished DRC-Hubo from every other humanoid in the competition. The robot could kneel and roll on small motorised wheels mounted just above its knees, switching back and forth between bipedal walking and a rolling kneeling stance.[13] KAIST used the rolling mode wherever the terrain allowed, including for the long approach between tasks and for the final dash to the stairs, slashing transit time and almost eliminating the risk of falling.[20] Many other teams burned the bulk of their hour on careful, slow, statically stable walking; DRC-Hubo simply rolled.[20]
IHMC's Running Man performed the closest thing to a perfect run by an Atlas. Running Man scored all 8 points and only stumbled at the very end, falling over while waving to the crowd after the buzzer.[22] Tartan Rescue's CHIMP, a four-limbed robot with track-equipped feet, delivered a memorable Day 1 recovery: the robot fell at the door, pushed itself back upright using its arms in roughly 30 minutes of carefully supervised motion, then went on to complete the remaining tasks.[17] CHIMP became one of the few robots in the entire event to recover from a fall on its own.[22]
The Finals are remembered as much for what went wrong as for what went right. Across the two days, viewers watched dozens of falls. Many were dramatic.[21] The IHMC Atlas tipped backward at the door on Day 1 and lay still for the rest of the run. The MIT Atlas pitched sideways during the egress on Day 2. The Korean DRC team RoboSimian slipped during the rubble task. Team THOR's robot collapsed within seconds of the start of its second run. Team TRACLabs's Atlas, having already cleared egress, lost its balance just past the door and sprayed roughly two metres of pink hydraulic fluid across the ground before workers carried it off.[21] The IEEE Spectrum compilation video of those falls, edited by Erico Guizzo and Evan Ackerman, became one of the most-watched robotics clips of the decade and is often used in lectures to illustrate why bipedal balance is hard.[21]
The sheer volume of failure data turned the Finals into one of the most influential robotics datasets ever generated. Researchers wrote dozens of post-mortem papers on what had gone wrong, including the widely cited Carnegie Mellon report "What Happened at the DARPA Robotics Challenge, and Why" by Christopher Atkeson, Benjamin Stephens, Jerry Pratt, and collaborators.[11] Recurrent themes in those analyses included unmodelled compliance in hydraulic actuators, the cost of operator decisions made under time pressure with degraded telemetry, and the brittleness of behaviours that had been hand-tuned in low-fidelity simulation.[11]
A wide variety of robot architectures appeared at the Finals, and the diversity is itself an important record of where humanoid and quasi-humanoid design was in 2015. The table below summarises the most prominent platforms.
| Robot | Team(s) | Class | Height | Mass | Actuation | Notable Feature |
|---|---|---|---|---|---|---|
| Atlas (DRC version) | IHMC, MIT, WPI-CMU, TRACLabs, VIGIR, TROOPER | Bipedal humanoid | 1.88 m | 156 kg | Hydraulic | Standard Track B reference platform built by Boston Dynamics |
| DRC-Hubo | Team KAIST, DRC-Hubo at UNLV | Bipedal humanoid | 1.75 m | 80 kg | Electric | Knee wheels for biped-to-wheeled transformation |
| CHIMP | Tartan Rescue (CMU NREC) | Four-limbed | 1.5 m | 200 kg | Electric | Tracked feet on each limb for stable rolling |
| RoboSimian | NASA JPL | Four-limbed ape-like | 1.4 m | 110 kg | Electric | Symmetric 7-DoF limbs with end-effector wheels |
| Running Man (Atlas) | IHMC | Bipedal humanoid | 1.88 m | 175 kg (post-upgrade) | Hydraulic | Best-performing Atlas in the field |
| Momaro | NimbRo Rescue (Bonn) | Four-wheeled centaur | 1.35 m | 60 kg | Electric | Steerable wheel feet plus dual manipulator arms |
| HRP-2 Kai | AIST-NEDO | Bipedal humanoid | 1.71 m | 65 kg | Electric | Updated AIST research platform from Japan |
| WALK-MAN | Team IIT WALK-MAN | Bipedal humanoid | 1.85 m | 132 kg | Electric (series elastic) | IIT Genova humanoid with compliant joints |
| THORMANG / THOR | Team THOR, Team ROBOTIS | Bipedal humanoid | 1.45 m | 47 kg | Electric | Virginia Tech and ROBOTIS collaboration |
| HRP-2 + Schaft predecessor | NEDO-JSK, NEDO-Hydra | Bipedal humanoid | 1.55 to 1.71 m | 50 to 65 kg | Electric / hydraulic | Tokyo and AIST-led entries |
| ESCHER (Team VALOR) | Team VALOR (Virginia Tech) | Bipedal humanoid | 1.7 m | 77 kg | Electric (series elastic) | Compliant whole-body humanoid |
| Aero | Team AERO | Bipedal humanoid | 1.7 m | 70 kg | Electric | Custom South Korean humanoid |
A few patterns are visible in this list. The hydraulic Atlas dominated the Track B contingent because of its raw power and the standardised software ecosystem around it. Electric humanoids designed in Korea and Japan, especially DRC-Hubo, were lighter, less prone to leaks, and easier to keep running across long competition days; KAIST took advantage of those properties.[13] Quadruped-style hybrids like CHIMP, RoboSimian, and Momaro skipped bipedal walking entirely for stability, trading the visual appeal of a humanoid for the practical advantage of always having a wide base of support.[15] None of the bipedal robots in the field could reliably stand back up after a fall under the time pressure of a single one-hour run; the four-limbed designs could in some cases self-recover.
The Finals carried a 3.5 million dollar total prize purse. The grand prize was 2 million dollars to Team KAIST. IHMC Robotics received 1 million dollars for second place, and Tartan Rescue received 500,000 dollars for third.[23] Several non-cash prizes also circulated through the event, including the Atlas hardware contracts that DARPA provided to the top VRC finishers and the simulation runtime that DARPA underwrote on Amazon Web Services. The Boston Dynamics Atlas hardware loans alone were valued at roughly 2 million dollars per unit at the time of delivery.
One of the most discussed elements of the DRC was the deliberate degradation of operator-to-robot communications. DARPA published detailed link specifications well in advance so that teams could test their software against realistic constraints.[38] The structure was:
Link 3 was deliberately too slow to permit traditional joystick teleoperation. Operators could send keyframes, target poses, gripper commands, and high-level skill invocations, but they could not stream continuous joint commands. The intent was to force teams to develop what DARPA called supervised autonomy: a model where a human supervises a partially autonomous robot through a constrained interface, shifting work between human and machine as the situation demands.[6] Most successful teams converged on graphical operator interfaces that let a supervisor click on a 3D point cloud, define a manipulation goal, and let onboard planners and controllers fill in the joint-level details.[12]
The DRC's lasting software contribution was the maturation of the Gazebo simulator and the surrounding open-source robotics tooling.[37] Through the OSRF contract, Gazebo gained a large number of features that have since become standard, including improved contact dynamics, multi-camera and depth sensor simulation, plugin-based ROS integration, and cloud deployment.[26] The drcsim package, which packaged a high-fidelity Atlas model and the DRC task environments, was open-sourced and remains a reference test bed for whole-body humanoid control.[27]
Many of the software stacks built for DRC teams matured into widely used libraries. IHMC's open-source whole-body controller for Atlas, written in Java and Scala, became a reference implementation of capture-point and momentum-based balance control.[12] Drake, the MATLAB and C++ optimisation toolbox developed initially by Russ Tedrake's group at MIT for the DRC, evolved into a major modern robotics library with applications well beyond humanoids.[16] The Robot Operating System (ROS) saw heavy investment during the DRC era and emerged as the de facto research middleware.[26] Conversely, the DRC also exposed the limits of the contemporary robotics software stack: the time required to debug perception failures over a degraded link, the difficulty of testing fall-recovery behaviours safely, and the brittleness of hand-tuned controllers all motivated the research push toward learned policies that gathered force after 2017.[11]
The DRC ended on June 6, 2015, but its influence on robotics in the years that followed is hard to overstate.[22] A generation of researchers and engineers who had spent two or three years living and breathing humanoid disaster response became the technical backbone of a wave of new robotics companies. The most direct lineages include:
The DRC is widely credited as the single largest catalyst for the surge of humanoid robotics startups that followed in the late 2010s and early 2020s.[9] Companies such as Apptronik (founded 2016), Agility Robotics (founded 2015 by Oregon State alumni who had been adjacent to the DRC community), 1X Technologies, Sanctuary AI, Figure (2022), and Tesla's Optimus program all build on the controls and perception research that the DRC funded and forced into reproducible, contestable form. When Boston Dynamics introduced the all-electric Atlas in 2024 and 1X demonstrated the NEO Beta humanoid the same year, both leaned on whole-body control and footstep planning ideas that had been pioneered at the DRC.
Not all post-DRC reactions were celebratory. A widely circulated 2015 Popular Science article called the Finals a bust, arguing that the public had expected too much and that the long, slow runs and frequent falls demonstrated how far the field still had to go. That criticism was, in retrospect, both fair and incomplete. Fair, because no humanoid robot in 2015 was anywhere near deployable in a real disaster. Incomplete, because the slow, falling humanoids of the Finals turned out to be the seedbed of the next decade of progress, and the failure data that the event generated was a gift to subsequent researchers.[22]
IEEE Spectrum dedicated extensive coverage to both the Trials and the Finals, with Evan Ackerman and Erico Guizzo producing daily reports, video compilations, and post-event analyses.[22] The Finals generated a special issue of the Journal of Field Robotics (volume 34, number 2, March-April 2017) titled "The DARPA Robotics Challenge Finals: Results and Perspectives", with a lead overview paper by Eric Krotkov and DARPA staff and contributed papers from most of the top finishing teams.[7] The same year, Springer published an edited volume of the same title that collected expanded write-ups from KAIST, IHMC, Tartan Rescue, NASA JPL, MIT, WPI-CMU, NimbRo, IIT, and others.[8]
Program manager Gill Pratt published a widely cited 2015 paper in the Journal of Economic Perspectives titled "Is a Cambrian Explosion Coming for Robotics?" that drew on the DRC experience to argue that several converging trends were positioning robotics for a sustained burst of capability growth.[9] In it, Pratt wrote that "two newly blossoming technologies, 'Cloud Robotics' and 'Deep Learning,' could leverage these base technologies in a virtuous cycle of explosive growth."[9] Pratt left DARPA shortly after the Finals to head the Toyota Research Institute, where his bets on data-driven robotics carried the DRC ethos into an industrial setting.
Workshops at major conferences continued to refer to DRC results for years afterward. The Robotics: Science and Systems (RSS) conference hosted multiple post-DRC workshops on humanoid balance, supervised autonomy, and benchmarking. The IEEE-RAS International Conference on Humanoid Robots (Humanoids) became the long-running venue where DRC alumni teams continued to publish their evolving systems.[10] The annual Conference on Robot Learning, founded in 2017, picked up many of the threads about learning-based control that the DRC had highlighted as needed but had not yet enabled.
The DRC was not without criticism, both during and after the event. The most commonly raised concerns were:
Despite these criticisms, almost every retrospective converges on the same conclusion: the DRC was money well spent. It demonstrated that the field could organise around a hard problem, attracted hundreds of millions of dollars of follow-on private investment, trained a generation of robotics engineers, and generated the simulation infrastructure that supports much of the contemporary humanoid research community.[1]
| Date | Event |
|---|---|
| March 2011 | Fukushima Daiichi nuclear disaster occurs in Japan. |
| Late 2011 | DARPA staff begin internal planning for a disaster-response robot challenge. |
| April 10, 2012 | DRC publicly announced. |
| October 2012 | Formal program kickoff meeting held. |
| June 17 to 21, 2013 | Virtual Robotics Challenge runs in the cloud using Gazebo and Atlas in simulation. |
| August 2013 | Boston Dynamics Atlas humanoid publicly unveiled by DARPA. |
| December 20 to 21, 2013 | DRC Trials held at Homestead-Miami Speedway. SCHAFT wins. |
| Late December 2013 | Google acquires SCHAFT and Boston Dynamics. |
| Early 2014 | DARPA announces SCHAFT will withdraw from the funded portion of the program. |
| 2014 | DRC Finals postponed from late 2014 to mid 2015 to give teams more development time. |
| June 5 to 6, 2015 | DRC Finals held at the Fairplex in Pomona, California. Team KAIST wins. |
| 2015 to 2017 | Special issue of Journal of Field Robotics and Springer volume publish the technical results. |
| November 2018 | Alphabet shuts down SCHAFT after failing to find a buyer. |
| 2017 to present | DRC alumni teams found, join, or anchor most major humanoid robotics startups, including Apptronik, Agility Robotics, Figure, 1X, Boston Dynamics' new Atlas program, Rainbow Robotics, and Tesla Optimus. |