Robotics Summit panel explores the state of humanoid robot design

Whereas robotic arms have arguably been mastered for manufacturing, it’s one other factor fully to design and construct a bipedal robotic that may stroll and manipulate objects. Plus, there’s the added complexity of that system working in a fast-paced surroundings with human staff, forklifts, and different equipment. Finally month’s Robotics Summit & Expo, a keynote panel targeted on the state of humanoid robots.

The session boasted a star-studded lineup:

• Al Makke, head of humanoid robotics for North America at Schaeffler
• Mike Nielsen, chief advertising officer at RealSense
• Aaron Prather, director of the Robotics & Autonomous Programs Program at ASTM Worldwide
• Alberto Rodriguez, director of robotic habits for Atlas at Boston Dynamics
• Pras Velagapudi, chief know-how officer at Agility
• Mike Oitzman, moderator and senior editor at The Robotic Report and Automated Warehouse

The Robotics Summit & Expo, held at Boston’s Thomas M. Menino Conference & Exhibition Heart, included unimaginable keynotes, displays, and panel discussions in regards to the world of robots, from their design to their implementation. Offered by The Robotic Report, roughly 3,900 attendees crammed the displays and explored the exhibit corridor, seeing every little thing from part producers to tennis ball-shooting robots.

Submit your session concept for the 2026 RoboBusiness

Humanoid builders look past the demos

For Boston Dynamics, its North Star for its Atlas program has been enabling it to perform as a general-purpose machine for bodily labor. Rodriguez stated that one of many issues the corporate has seen whereas deep diving with many shoppers is that, aside from only a few functions with very giant and secure scale, the norm is that the majority jobs are one among a form.

“Our roadmap is constructing the know-how that’s crucial to advertise that general-purpose case on the ranges of {hardware}, on the ranges of the fashions and architectures which are driving the habits, and likewise very importantly on the degree of the deployment technique,” he stated. “For the mixing technique, if you happen to fail on discovering a basic technique for any a kind of three issues, it turns into too costly.”

Rodriguez defined that Boston Dynamics has began with logistics in manufacturing, an software the place the corporate and its mum or dad Hyundai suppose there’s a good stability between generality and complexity.

“You must deal with all of the elements — for instance, like going to a cart — nevertheless it’s nonetheless not shut sufficient to the meeting line that you must cope with the timing constraints and the security constraint of getting to work proper subsequent to different individuals,” he famous. “Final yr, we introduced Atlas as a form of first train right into a manufacturing unit to do a primary proof-of-concept demonstration of absolutely data-driven structure driving habits and sequencing state of affairs.”

“We introduced it to CES this January for an entire week, simply demoing, and subsequent yr, we’re going again,” Rodriguez added. “This yr, we’re going again to the manufacturing unit to point out a extra end-to-end demonstration of Atlas — a full studying pipeline, not simply dealing with the habits, however dealing with your entire workflow, connecting to the manufacturing unit, and dealing with exceptions.”

He additionally famous that in Boston Dynamics’ journey to mass-production scale, the corporate has now secured sufficient prospects (together with Hyundai) that it has dedicated to deploy on the order of 25,000 humanoids in factories. Boston Dynamics has made an extra dedication to ramp up manufacturing capability to 30,000 Atlas robots per yr by 2028.

Agility (previously Agility Robotics) has additionally moved previous pilot initiatives. It has labored on Digit humanoid deployments with corporations together with Amazon, GXO, Schaeffler, Toyota, and Mercado Libre.

“We’ve actually been increasing the industrial facet and studying from that to determine — what do we have to shut the remaining gaps that enable us to scale?” stated Velagapudi. “That’s a few various things: It’s discoveries round what we would have liked out of the security case, which has been an enormous piece of all of this with an extremely highly effective, dynamically secure robotic. How can we transfer round these amenities and have the ability to work in nearer proximity or with out guarding round people?”

Velagapudi additionally famous that Agility has spun up an ISO committee/working group with Boston Dynamics and others to check the security situation and give you an answer that it’s going to incorporate into the subsequent era of its robotic.

He added that Agility has been increasing past its preliminary functions of tote, case, and container manipulation. It’s now transferring towards merchandise manipulation within the subsequent yr or so.

Robotics security and requirements efforts proceed

The state of requirements in robotic security worldwide is evolving as quick because the robots themselves are. Prather defined that ASTM is conscious of and concerned in quite a few initiatives.

“ISO has two new teams: security chaired by Boston Dynamics, after which China is main the information effort,” he stated. “That is the subsequent effort that we’re working with NIST on … a proposal for a humanoid check mattress of about 10 exams.”

Prather detailed how the check beds would come with capabilities comparable to locomotion and manipulation. Some producers could have these exams shipped to them, to allow them to use them internally. However one other purpose is to make use of these check beds for future competitions, so groups can truly put their robotic to the check in entrance of the world at robotics occasions.

These exams are additionally meant to assist the event of robotics requirements, famous Prather. “Security efforts are beneath method at ISO, ASTM, and NIST,” he stated. “We’re beginning to work on the efficiency repeatability exams, and there are quite a few different efforts happening.”

In response to a query about humanoid security, Prather stated he hoped that the primary security requirements shall be drafted throughout the subsequent two years.

Nevertheless, AI basically modifications the panorama, acknowledged the panelists. Conventional security methodologies which are deterministic are much less fitted to advanced methods that contain leaned behaviors. Among the new engineering challenges right here embrace:

• Precisely predicting failure habits
• Repeatability of efficiency
• Threat evaluation of human-robot interactions
• Validation of AI-based determination making

Rendering of a proposed equipment for standardized testing of humanoid robotic capabilities. Supply: NIST, ChatGPT

Notion points to be overcome

Nielsen informed the gang on the Robotics Summit that RealSense’s progress has been “extraordinary” and that issues are transferring in a short time.

“From a development perspective, we’re seeing just a few issues concurrently, and we’re capable of do issues quite a bit sooner now,” he stated. “A kind of is partnerships with different tech distributors in each the silicon and the broader AI area, so we’re doing quite a lot of work proper now round simulation. The sim2real hole is closing.”

“It’s been improbable work, however there’s quite a lot of work to do,” stated Nielsen. “There’s some actually fascinating exercise proper now with NVIDIA and some others on growing … a common imaginative and prescient mannequin, which lets you take a RealSense digicam in any modality that you just need to and making use of these in Isaac, so you may land it in actual time, together with noise fashions.”

“The most important hole that we see is taking a pure mannequin and making use of that in the true world and ranging from scratch,” he famous. “As a result of if you happen to don’t have real-world implications constructed into the mannequin, you then’re actually simply utilizing face and eyeballs.”

Nielsen additionally stated his firm is studying quite a bit from China, the place the shoppers are transferring extraordinarily quick. He defined that the product life cycles in China iterations are on the order of 30 to 40 instances sooner than they see in different areas, as a result of the tolerance for threat is greater.

“We’re seeing … envelopes being pushed from a mechatronics perspective, issues like dancing robots, together with what we noticed on SportsCenter the opposite night time,” he stated. “That transition from robots that may transfer slowly to robots that may transfer and do judo then has the subsequent step to it, which is how they do issues autonomously, and [does] the autonomy stack?”

In accordance with Nielsen, the notion necessities for humanoid robots are very completely different than for different methods.

“There’s some fundamental modalities the place you want density of data at close to pace or at close to vary,” he stated. “The speeds are completely different, they usually differ based mostly on the z from the robotic, as a result of a swinging arm is quite a bit sooner than a turning torso.”

“Issues that fall are very completely different than issues as an entire, so we truly had an incident the place one among our engineers was doing one thing he in all probability shouldn’t have completed with a robotic with a buyer, and it fell on him,” stated Nielsen. “So, these are very completely different failures we’re coping with which have an enormous reliance on notion — to determine once they fail, along with the speeds, the distances, the density of data, after which the scene reconstruction.”

Rodriguez additionally raised an fascinating level regarding downscaling higher-resolution imagery right into a dimension extra appropriate with what neural networks can deal with.

“One of many issues that we’ve seen that’s been a dramatic change is that we get far more worth from much less pixels which are greater high quality, however finally the habits architectures which are driving most of what humanoids do at present find yourself downscaling photos to love 240 by 240 pixels, as a result of that’s what we feed into neural networks,” he stated. “We can not feed large photos [into the neural networks], so finally what occurs is that we take away quite a lot of info that has been captured.”

“As a substitute, we’d be higher off with higher-quality pixels which have greater dynamic vary which have a world shutter, for instance — that don’t have issues when the torso is transferring quick, which have excessive decision in proximity, or focus in proximity,” stated Rodriguez.

The $20,000 humanoid query

One of many business’s extra enduring questions revolves across the economics of humanoids: Can the price of the know-how ever come right down to the place a $20,000 value level is achievable? The panelists expressed cautious optimism in regards to the idea, because of automotive-style manufacturing strategies, rising quantity of manufacturing, consolidation of sensors, higher actuator design, and some extent of part standardization.

“There are very fascinating pressures, clearly from some individuals on the stage, in addition to from a bunch of different prospects,” stated Nielsen. “The sensor-stack prices may be actually excessive, particularly on AMRs, the place it will get like tremendous harmful, as a result of they’re carrying heavy stuff.”

“The whole notion stack of a robotic is about $20,000, practically the value of what we’re speaking about for your entire robotic. It’s basically untenable,” he added. “So, among the work we’ve been doing is in collapsing the modalities. Generally you want sparse info, and generally you may have to have the ability to learn a bar code. These are basically completely different issues to unravel for, however you don’t need to purchase actually two completely different cameras to unravel for these two sorts of issues. In order that analysis is being accelerated at RealSense — fewer sensors is sweet, with out dropping 360-degree notion.”

Makke famous that on the 2025 Robotics Summit & Expo, a panel member stated that they have been somewhat bit pessimistic in regards to the idea.

“I stated, ‘I see your pessimism, and I elevate you with optimism,’ and I nonetheless really feel the identical method, in all probability much more so,” Makke joked. “One factor has to occur, and it’s a hen or the egg scenario. You want a sure invoice of fabric [BOM] price to justify an ROI, so the quantity will increase.”

“You must check out these low TRL [technology readiness levels] applied sciences, and you must be prepared to fail somewhat bit,” he stated. “Now we have to commoditize among the key parts. Actuators are an enormous a part of the BOM prices, motors are as effectively, and sensors and PCBs are as effectively.”

Makke stated that if he needed to guess on a timeline for that form of pricing, he would say within the subsequent three to 5 years.

The submit Robotics Summit panel explores the state of humanoid robotic design appeared first on The Robotic Report.