Dennis Hong first spied Japan’s ruined nuclear power plant from a bus wrapped in plastic. A hefty layer of protection guarded the seats, floors and handles from radioactive dust. Hong wore a face mask and gloves to limit his exposure. Like the other passengers, he had dressed in old clothes that he was willing to toss after the trip.

More than three years earlier, after an earthquake and tsunami battered Japan’s eastern coast, portions of the Fukushima Daiichi power station blew, blasting radiation into the sea and sky. Today, villages outside the plant still lie as barren as ghost towns. Along the coast, smashed buildings, flipped cars and train tracks twisted like taffy stand as reminders of the catastrophe.

“It’s like a disaster site frozen in time,” Hong says. “It’s surreal.”

Workers toiled day and night to save the plant, but they had to get out as radiation levels rose. Even today in disaster areas not tainted with radiation, picking through the destroyed buildings is treacherous: People need to dodge shards of glass and metal and duck clouds of smoke and dust.

Ideally, robots could take over for human crews. But seemingly simple tasks, such as moving, communicating and staying powered up, pose big challenges for machines.

THOR-OP, or Tactical Hazardous Operations Robot-Open Platform. (Dennis Hong/Robotics & Mechanisms Laboratory (RoMeLa)/UCLA)

Hong, a UCLA roboticist, is one of several engineers working to make robots that can come to the rescue in disasters. He and others from academia, industry, NASA’s Jet Propulsion Lab and the Defense Department’s research agency DARPA traveled to Fukushima last spring to see what they were up against.

“The take-home message was ‘Wow, it’s damn difficult,’ ” Hong says.

A better bot

Engineers have built impressive- ­looking humanlike bots that can play trumpet and even compete against each other in slow-moving soccer games. But machines that can actually do the work of humans in disaster zones — climbing over rubble, digging through debris for survivors, opening doors and valves — don’t exist.

So DARPA kicked off a contest to create robots that someday could do the job. In 2012, the agency announced a competition designed to push disaster robotics technology miles past where it is today. A year ago, 17 robotic contenders, including Hong’s human-shaped machine, THOR-OP, tackled a rugged obstacle course to try to gain a spot in the finals, to be held in June 2015.

Gill Pratt, a DARPA program manager, knows that it may take years to develop robots that might have saved the power plant. But he thinks the competition — with its motley crew of robotics engineers and their rowdy fans — is a good place to start.

9/11 lessons

In the past decade or so, disaster- response robots haven’t changed much. When the World Trade Center towers were destroyed in 2001, engineers deployed a handful of lightweight bots to burrow through the rubble.

“These guys went into spaces where the first responders couldn’t go,” remembers Robin Murphy, a field roboticist at Texas A&M University. Because 110 floors of concrete and steel collapsed into dense piles of debris, she says, ground- ­penetrating radar couldn’t see through the rubble and search-and-rescue dogs had trouble sniffing out victims.

The robots, roughly the size of shoe boxes, offered a new way for rescue teams to take a look. Murphy considers the roving machines a success: A few tunneled deep within the wreckage and withstood extreme heat to find 10 sets of human remains. But the robots didn’t locate any survivors, and they ran into a slew of technical snags.

One robot slipped its tread and had to be pulled from the rubble and repaired. Another got wedged in a gap, stalled until a crew could tug it out by its safety tether. A third lost communication, broke loose from its tether and was never seen again.

Machines at Fukushima faced similar problems. In the days following the earthquake, researchers from around the world sent ground-based and aerial robots to Japan to try to reach spots where humans couldn’t safely travel. DARPA sent all it had, Pratt remembers. But training power plant personnel to use the robots took time. And once inside the plant, muddy steps, tight stair landings and a severed communication cable thwarted the machines’ progress.

“We did everything we could, and it still wasn’t good enough,” Pratt says.

A far cry from C-3PO

Years before Fukushima, Japan — a country known for its advanced robotics — had actually developed robots to respond to nuclear emergencies.

After a power plant accident in 1999, engineers created six huge, treaded robots to open doors, turn valves, carry heavy loads and even clean up radiation. But these bots weren’t maintained well enough for use at Fukushima. Even if they had been ready to go, they may have been too big and heavy — more than 440 pounds — to be useful, engineer Keiji Nagatani and colleagues reported in the Journal of Field Robotics in 2013.

A smaller robot, designed for disposing of explosives, was the first to enter Fukushima’s reactors, about a month after the disaster. Shin-height and light enough to be carried on a soldier’s back, the robot, called the PackBot, uses treads to haul itself over bumpy turf. A skinny arm mounted with a gripper claw and a camera lets the robot grasp and see. In the days following the Fukushima earthquake, iRobot, the Massachusetts company that produces the PackBot, rushed to add radiation and chemical sensors to the bots.

The sensors came in handy: Plant workers used them to find places inside the plant where humans could safely explore. But maneuvering the robots through the darkened plant was at times impossible. Though the machines could open doors, the bots’ operators had to use two PackBots to do so: one to turn the handle and the second to aim its camera at the first. What’s more, the robots struggled to climb the plant’s slick metal stairs.

After the PackBot failed to make it up a few flights, workers sent in Quince, Japan’s version of the machine. But Quince couldn’t adapt to stairway landings that were smaller than expected, Murphy explains, and the robot got stuck. Quince also snapped its tether, leaving the bot stranded in a place where humans couldn’t save it.

Gizmos in the lab

Japan’s lack of versatile, usable rescue robots surprised people in the field, says University of Pennsylvania roboticist Mark Yim. Not only did Japan have simple ground-roving bots such as Quince, the country also had more humanoid machines than any other country, he says. Unfortunately, these fledgling gizmos weren’t ready to leave the lab. For some such devices, just making it across slightly uneven floors can be tricky.

“People see a humanoid robot and they think it can do anything a human can do,” Yim says. “In reality, it can do very little.”

Still, some type of limbed, humanlike machine might be the key to tackling disasters in urban areas, where stairs and doors have been designed for human legs to climb and human hands to open.

“At a disaster site, there are bulldozers, excavators, power tools — all this great equipment that people use for rescue missions,” Hong says. “So, naturally, a lot of people believe the robot needs to be in a human form.”

And in the DARPA Robotics Challenge trials, most contenders were. But these robots aren’t anywhere near the C-3PO or Terminator-style machines of the movies. Today’s state-of-the-art humanoids have just begun venturing out of the lab, and they’re taking baby steps.

Clash of the bots

Watching disaster-relief robots triumph — or fizzle — at even the simplest tasks can be thrilling. Last December, thousands of cheering spectators flocked to a Florida racetrack to see a medley of high-tech machines in action at the trials. The robots lumbered and lurched through a series of tasks, including opening doors, traveling over rubble and turning valves. Here was the future of disaster-relief robots, where the simple act of getting a bot to set a tool on a table is still considered showboating.

“A lot of these are tasks that a person could complete in a minute or less,” says software engineer Doug Stephen of the Florida Institute for Human and Machine Cognition in Pensacola. “But things we take for granted are incredibly hard for a robot.” And figuring out how to build a bot that can sail through all of these tasks, rather than designing a specialized door-opener or valve-turner, is even harder.

The competition, however, may have already begun to pay off. Teams at the trials dived into the problems that plague humanoid robots — walking, power usage, handling tools. They’ve created a bevy of bots that could one day do some good in a disaster.

Google snapped up the Japanese company that built the trials’ first-place finisher, SCHAFT, which scored an impressive 27 out of 32 points at last year’s trials.

The second-place finisher, with 20 points, may be on its way to overcoming a huge challenge in bipedal robotics: locomotion. “Historically,” Murphy says, “a lot of focus has been on just walking and not falling over.”

Humans don’t simply pick up and set down their feet, Stephen says. People tailor their steps on the fly, weaving through busy streets, striding over sidewalk cracks and swiftly rebalancing themselves after stumbles. And they make most of these adjustments subconsciously.

‘No brains of its own’

“But the robot has no brains of its own, so we have to tell it how to do things,” Stephen says. Capturing the hair-trigger tweaks and easy grace of human gait “is a tough nut to crack.” For the DARPA challenge, he and colleagues designed software that helps robots “think” on their feet.

Working with a heavy-duty robotic hulk named Atlas, which was developed by Boston Dynamics, the Pensacola team cruised through the rough terrain task at the trials. The monster of a machine, bulging with black wires and a silvery skeleton, stands about as tall as Arnold Schwarzenegger and weighs more than the action hero did in his bodybuilding days. “Atlas is designed to take a beating,” Stephen says. “It basically looks like a giant roll cage.”

In the trials, Atlas conquered a task that involved hiking up a ramp, climbing steps and navigating over jumbled cinder blocks. A large crowd cheered as the robot picked its way through the course — step by ginger step — without falling.

Other teams handled the walking problem in a less humanlike way. Engineer Brett Kennedy and colleagues at NASA JPL in Pasadena, Calif., built and developed software for a bot they named RoboSimian. Short and sleek, with a headless body and four identical limbs, the lithe little robot is Atlas’s opposite. Instead of walking on two legs, RoboSimian scoots around on two wheels attached to its rear. Or it crawls on all fours.

Bypassing the need to balance on two feet saves RoboSimian power. In a disaster, the bot might be able to park itself on its limbs, set the brakes and subsist on tiny sips of energy from a battery while waiting for orders from humans. “It’s a very patient robot,” Kennedy says.

Nonhuman ways

Finding nonhuman ways to tackle problems might even give humanoids the best of both worlds. Dennis Hong’s robot, THOR-OP, which looks something like a leaner, friendlier Atlas, was able to turn a valve in one of the trials’ tasks by spinning its wrist rather than turning the wheel hand over hand, as a human would.

Hong’s team placed ninth, just shy of advancing to the finals. But then THOR-OP made the cut when SCHAFT dropped out of the competition. DARPA expects additional teams to compete in the finals, including some sponsored by the European Union, Japan and Korea.

Though the agency hasn’t released many details about the upcoming finals, the teams know that their bots will have to operate without a power cord, and the communication link between robots and team members will switch on and off, like a cellphone dropping a call, Pratt says. Instead of attempting (or opting out of) individual tasks, the bots must power through several tasks in just one hour. The winning team receives a $2 million prize.

Engineers competing in the finals agree that even the best bots of the competition probably wouldn’t be ready for another Fukushima. But, Kennedy says, “I think that we would be able to do better.” And the range of technology the competition inspires may lead to many different types of disaster-ready robots.

“Someday, it would be really cool to have a robot that does everything a fireman can do,” Kennedy says, “burst through the door, save the baby, the whole nine yards.”

DARPA Robotics Challenge tasks: In the 2015 finals, robotic contenders will navigate through a gantlet of tricky tasks. DARPA hasn’t finalized the course yet, but it may include a combination of the following directives:

●Drive a vehicle.

●Get out of the vehicle.

●Open a door and travel through the opening.

●Move debris or climb over it.

●Use a tool to cut a hole in a wall.

●Reach through an opening and open a valve.

●Cross over a field of loose debris and pipes.

●Insert a cylindrical plug into a receptacle.

●Climb stairs with a rail on one side.

●Execute a surprise manipulation task.

A version of this article appeared in the Dec. 13 issue of Science News and can be read online.