How AI Is Rewriting the Timeline for Useful Robots

For a long time, the central challenge in robotics was teaching machines exactly what to do. Now, some of the most interesting advances are coming from robots that can learn what to do for themselves.

That shift is changing expectations for how quickly robots could become genuinely useful. New systems can learn from demonstrations, adapt to unfamiliar situations and transfer what they’ve learned across tasks and machines—bringing robotics closer to the flexibility that made generative AI so disruptive.

Members of the Senior Executive AI Think Tank—a curated group of experts specializing in machine learning, generative AI and enterprise AI applications—have been watching these developments from across technology, healthcare, infrastructure, transformation and enterprise AI. We asked them what development in AI robotics has been most interesting to them this year, what it changed their minds about and what it suggests about how quickly robots will become capable of doing more in the real world.

Learn to Fail, Then Improve

For Andre Shojaie, Founder of HumanLearn, the most important robotics development is not a dramatic physical feat. It is the ability to learn from failure.

“What’s changing is the ability of robots to encounter failure, interpret what went wrong, adapt and try again with far less human intervention,” he says. “A robot doesn’t need to be perfect on day one if it can improve safely in the environment where it works.”

This matters because traditional robotics economics often assume substantial engineering effort before deployment. If a system can safely improve after deployment, some of that upfront cost can move into the operating life of the robot.

For executives, Shojaie’s insight suggests evaluating robots less like finished equipment and more like continuously improving software products—with explicit limits around what the system can learn autonomously.

“The breakthrough may not be the moment robots can do everything humans can,” Shojaie says. “It may be the moment deploying an imperfect robot becomes economically rational because tomorrow’s robot is better than today’s.”

Make Skills Portable Across Robot Bodies

Dr. Aditya Vikram Kashyap, Vice President of Firmwide Innovation at Morgan Stanley, sees portability as one of the year’s most important robotics advances.

He points to Gemini Robotics 2’s ability to adapt across robot bodies. Google DeepMind reports that its on-device model can adapt to new bi-arm embodiments with a few hours of training and typically fewer than 200 examples.

“The implication is that capability can become portable,” Kashyap says. “Progress on one platform need not stay trapped there.”

Instead of rebuilding intelligence for every new machine, organizations could increasingly reuse capabilities while performing relatively modest local adaptation.

But Kashyap sees an important distinction between capability and dependable deployment. 

“Dexterity in a demonstration is the beginning,” he says. “Commercial value requires robots that recover from mistakes, handle unfamiliar conditions and seek help before creating harm.”

For executives, that means tracking two metrics separately: how quickly a robot can learn a task and how reliably it can execute that task under real operating conditions.

“It’s now clear that robots are rapidly evolving from task-specific tools to collaborative partners in patient care.”

– Will Conaway, President of Tuxedo Cat Consulting

SHARE IT

Move From Task Tools to Care Partners

Will Conaway, President of Tuxedo Cat Consulting, says the biggest surprise has been the combination of robotics and large language models. 

“One of the most fascinating developments in AI robotics this year is integrating large language models into healthcare robots, enabling them to understand nuanced patient needs and communicate empathetically,” he says.

While he previously believed human-like interaction was years away, he notes that recent breakthroughs, such as a robot’s ability to adapt care plans based on real-time conversations, has changed his perspective.

“It’s now clear that robots are rapidly evolving from task-specific tools to collaborative partners in patient care,” he says.

He believes this suggests emotionally intelligent robots will be part of everyday healthcare sooner than expected, which could transform both patient interactions and clinical workflows.

Make Specialized Robots Work as a Team

For Rishi Katdare, Senior Technology Executive at Amazon Web Services (AWS), the most revealing robotics experiment happened at home.

Katdare bought robotic vacuums for different floors because the machines could not climb stairs. When he placed two on the same floor, they bumped into each other and then continued independently.

“That small moment exposed a larger limitation,” he says. “Each robot could complete a task, but neither could operate as part of a team by sharing context, dividing work or adapting to what the other had done.”

This is where he sees the next leap in household robotics: not better hardware, but when specialized robots can become coordinated agents with a shared goal.

“When machines can delegate, coordinate and learn together,” Katdare says, “isolated automation becomes an operating system for the home.”

Watch the Organizational Clock

Divya Parekh, Founder of executive coaching brand DivyaParekh.com, says the important shift is from asking whether a robot can perform a task to asking whether it can learn the task from a single demonstration.

“If physical skills become promptable the way knowledge work did,” she says, “then capability moves at software speed, and the gating factor is no longer the hardware.”

But organizations still have to decide what they will permit physical AI to do. Parekh sees leaders moving through familiar stages of technology adoption, from fear and acceptance to confidence and agency.

“Fear still looks like resistance from the outside, and most organizations stall there,” she says. “So my timeline shortened for what robots can do and did not shorten at all for what organizations will let them do. The second clock is the one worth watching.”

For executives, that means an AI robotics strategy needs an adoption plan as much as a technology plan.

“Fleet learning and cross-embodiment training mean a skill learned on one machine can propagate to thousands of others, without a single lick of code.”

– John Cho, Chief Information Officer at EdgeConneX

SHARE IT

Rebuild the Stack Beneath the Model

John Cho, Chief Information Officer at EdgeConneX, observes that robotics intelligence is moving into learned models faster than the underlying physical and software infrastructure can adapt.

“We see robot behavior moved out of source code and into model weights,” Cho says. “Fleet learning and cross-embodiment training mean a skill learned on one machine can propagate to thousands of others, without a single lick of code.”

That creates enormous upside—but also exposes architectural weaknesses. Learned policies must ultimately interact with real-time control systems, operating systems, networks and physical actuators.

Cho is therefore “optimistic about capability but pessimistic about how deployable they really are.”

“Many in AI are just starting to realize the complex interplay between robotics software and hardware and how it is to execute against learned policies,” he says. “Real-time control moving to RTOS cores, micro-ROS, and the fight to replace DDS at scale are early signs the industry understands that the stack between the model and the motor cannot currently carry the weight(s).”

Teach Robots With a Demonstration

Rishi Kumar, Chief Transformation Officer at Matchingfit, points to Skild AI’s S1 as an early signal that physical skills can become “promptable.” The system is designed to execute previously unseen tasks from a single video demonstration, including long-horizon tasks lasting up to 10 minutes.

“The most consequential robotics development this year is robots beginning to learn new long-horizon tasks from a single human video,” Kumar says. “Physical skills may become ‘promptable,’ much like knowledge work became promptable with LLMs.”

That changes the scaling equation. Instead of every new task requiring hours of teleoperation and task-specific retraining, a reusable robot brain can potentially absorb demonstrations and generalize.

“The bottleneck is shifting from intelligence to trust,” Kumar says, pointing to recovery from edge cases, safety, economics and integration. “The strategic race is no longer robot versus human; it is learning velocity versus deployment friction.”

Turn Everyday Video Into Robot Training Data

Vivek Kumkar, Sr. GenAI Leader at Amazon Web Services (AWS), focuses on a resource robotics has historically lacked: scalable training data.

“The biggest robotics dataset already exists,” Kumkar says. “It is ordinary videos of people doing everyday tasks.”

The significance is not simply that robots can watch people. It is that existing video could become a source of reusable information about movement, intent and task structure without requiring every demonstration to be recorded on the exact robot hardware.

Kumkar says the bottleneck therefore moves toward data quality: “clear views of the hands, accurate labels for each action, and enough examples of things going wrong, not just the clean takes.”

He now expects warehouse and household manipulation to improve more quickly than hardware roadmaps suggest, noting that “the teams that get good at video data will get there first.”

Use Robots to Discover, Not Just Execute

For Sai Krishna Reddy Mudhiganti, Staff AI/ML Software Engineer at Samsung Semiconductor, the most interesting development is not a household robot at all. It is the rise of autonomous laboratory systems that use robots as part of an AI-driven discovery loop.

He points to the OpenAI-Ginkgo Bioworks autonomous lab experiment, in which GPT-5 designed experiments, a robotic lab executed them and the resulting data informed subsequent experiments. OpenAI reports that the system tested more than 36,000 reactions and achieved a 40% reduction in protein production cost in the experimental system.

“That changed my view of what makes a robot valuable,” Mudhiganti says. “It can help generate knowledge we don’t yet have.”

The implication is profound for R&D organizations. A robot does not have to resemble a human or replace a worker to create substantial value. It can compress the time between hypothesis, experiment and evidence.

“The milestone I’m watching is how cheaply robots can test an idea,” he says, “and how quickly that evidence leads to the next invention.”

“Handling a task once is impressive. Handling it safely when the object slips, the room changes or someone walks in is the real test.”

– Pon Murugesh Devendren, SAP Enterprise AI Architect at Deloitte

SHARE IT

Let Robots Feel Their Way Forward

Pon Murugesh Devendren, SAP Enterprise AI Architect at Deloitte, points to FTP-1, a generalist tactile policy designed to transfer learned manipulation skills across different tactile sensors and robot embodiments. The research trained on roughly 3,000 hours of data spanning 21 tactile sensors and found meaningful gains even on previously unseen sensor setups.

“What caught my attention this year was a robot learning to use touch, not just sight,” Devendren says. “If robots can carry that experience across different hands and machines, we may not have to teach every new setup from scratch.”

Vision can tell a robot where an object is, but touch can help it recognize whether an object is slipping or whether its grip is too strong.

“Handling a task once is impressive,” Devendren says. “Handling it safely when the object slips, the room changes or someone walks in is the real test.”

That is the kind of robustness executives should demand before expanding robotics beyond controlled environments.

What Leaders Should Take Away

• Design robots to learn safely from failure. The most valuable systems may be those that can improve after deployment without turning every mistake into an operational incident.
• Separate capability from dependability. A robot that can learn a task quickly still needs testing for unfamiliar conditions, recovery and safe escalation.
• Treat physical AI as a workflow technology. In healthcare and other high-touch environments, evaluate whether robots improve the broader experience rather than simply automate a physical task.
• Plan for robot-to-robot coordination. Specialized machines can become substantially more valuable when they share context and divide work around a common objective.
• Budget for organizational adoption. The technical timeline and the leadership timeline are different, so build governance, training and change management into the robotics roadmap.
• Modernize the stack beneath the model. Learned policies need deterministic controls, real-time infrastructure, cybersecurity and safety mechanisms to become deployable systems.
• Measure learning velocity. Track how much data, time and human intervention are required to teach a robot a new task.
• Build a video-data strategy. Existing footage can become a valuable source of training information if organizations invest in curation, labeling and failure examples.
• Look for discovery use cases. Autonomous laboratories show that robots can create value by accelerating experimentation and generating new knowledge, not just by performing manual work.
• Add touch and other physical signals to the AI stack. Robust manipulation will depend on systems that can react to what they physically encounter, not merely what cameras can see.

When Robots Learn, Organizations Must Adapt

The most interesting development in AI robotics this year isn’t necessarily a single robot. It is the emergence of a new learning loop: models learn from demonstrations, transfer skills across embodiments, improve from experience, use multimodal information and increasingly participate in systems where multiple machines or humans share a goal. The result is a shift from robots as programmed equipment toward robots as adaptable AI platforms.

That does not mean universal humanoid autonomy is around the corner. The Think Tank members point instead to a more practical—and potentially more important—future in which useful robots scale task by task, workplace by workplace and experiment by experiment. The capability clock is speeding up. For executives, the question now is whether their organizations can build the trust, infrastructure, data and operating models needed to keep pace.