In the world of robotics demonstrations, one task reigns supreme: folding t-shirts. You've likely seen the videos - sleek humanoid robots meticulously smoothing, flipping, and stacking laundry with an almost hypnotic precision. But why this obsession with apparel? The answer lies in what experts call the "optimal zone" of available technology.
Folding clothes is a task that appears deceptively complex for machines, involving soft, deformable materials, variable shapes, and real-time adjustments. Yet, with current AI methods like learning-from-demonstration - where robots are trained on repeated short sequences via puppeteering interfaces - companies can achieve high success rates, often above 80-90% in controlled settings.
It's flashy, feasible, and far from trivial, making it a perfect showcase for progress without exposing the deeper limitations of today's systems.
Benjie Holson, a veteran roboticist who spent eight years at Google X's Everyday Robots project and is now the founder of General Robots (as well as Director of Robotics at Robust AI), has been vocal about these constraints. In a September 2025 Substack post, he argued that while these techniques excel at laundry, they fall short of true generalization.
"It might feel like if our AI techniques can fold laundry maybe they can do anything, but that isn’t true, and we’re going to have to invent new techniques to be really general purpose and useful," Holson wrote. Current setups often lack wrist force feedback, advanced finger control (limited to open/close grips), tactile sensing, and high precision (typically 1-3 cm accuracy), confining robots to narrow, predictable tasks.
Holson's frustration peaked after viewing the inaugural World Humanoid Robot Games in Beijing, held in October 2025. Organized to spotlight China's AI prowess, the event featured over 500 humanoid robots from 280 teams across 16 countries, including the U.S., Germany, and Japan.
Competitions included soccer, boxing, running, and a 100-meter race, with robots demonstrating skills like scoring goals, recovering from falls, and even backflipping. The opening ceremony dazzled with robots performing hip-hop dances, martial arts, and playing instruments like keyboards and guitars.
Yet, despite the spectacle - one robot even modeled clothes alongside humans - Holson found it underwhelming, noting it didn't push boundaries toward practical, everyday utility.
Inspired (or perhaps provoked) by this, Holson proposed his own "Humanoid Olympic Games" as a benchmark for advancing humanoid robotics. The concept includes five disciplines, each with three escalating challenges: bronze (achievable with tweaks to current tech), silver (requiring moderate innovations), and gold (demanding breakthroughs in hardware or AI).
The disciplines focus on household chores to emphasize real-world applicability:
1. Full Body (Doors): Handling asymmetric forces and whole-body coordination.
- Bronze: Enter a round-knob push door.
- Silver: Enter a lever-handle self-closing push door.
- Gold: Enter a lever-handle self-closing pull door, blocking it with a limb.
2. Laundry: Building on folding demos but adding complexity.
- Bronze: Fold an inside-out T-shirt.
- Silver: Turn a sock inside-out.
- Gold: Hang a men's dress shirt (one sleeve inside-out) on a hanger, fixing the sleeve and buttoning at least one button.
3. Basic Tool Use: Emphasizing strength and dexterity.
- Bronze: Clean a window with Windex and paper towels (three spritzes, streak-free wipe).
- Silver: Make peanut butter sandwiches (scoop, spread, cut in half; jar starts/ends closed).
- Gold: Use a key from a keyring (insert and turn the correct one without dropping).
4. Finger Tips: Testing in-hand manipulation.
- Bronze: Roll matched socks.
- Silver: Use a dog poop bag (tear off, slide over hand, pick up mock waste, flip inside-out).
- Gold: Peel an orange without tools.
5. Slippery When Wet: Addressing wet, messy environments.
- Bronze: Wet a sponge and wipe a countertop.
- Silver: Clean peanut butter off the manipulator.
- Gold: Wash grease off a pan with water and a sponge.
Enter Physical Intelligence (PI), a San Francisco-based startup that's rapidly emerging as a leader in physical AI. In November 2025, PI announced they'd tackled Holson's challenges using their π 0.6 model - a vision-language-action (VLA) system fine-tuned on diverse datasets.
Surprisingly, they claimed gold in three disciplines: Full Body (opening and passing through a self-closing lever-handle door), Basic Tool Use (using a key, including reorienting it mid-air), and Slippery When Wet (cleaning a greasy pan with water and a sponge).
For the other two, they earned silver: turning a sock inside-out (Laundry) and handling a dog poop bag (Finger Tips).
PI's demos, available on their blog, include videos of these feats - such as the robot scrubbing a pan or flipping a poop bag inside-out (using mock waste, of course). However, success wasn't flawless: average task completion hit 52%, with 72% progress across attempts.
A baseline VLA without PI's pre-training managed only 9% progress and zero successes, highlighting their advancements. Crucially, most tasks required under nine hours of data collection, suggesting that as models scale, even complex skills could demand less data and tolerate noisier inputs.
The misses? Purely hardware-related. PI's simple, wide grippers couldn't fit into a dress shirt sleeve for the Laundry gold or peel an orange (Finger Tips gold), even with a makeshift tool (which violated rules).
"When we added a 'peeler' to our gripper, the task became feasible," they noted, but stuck to their ethos of versatile, low-cost hardware. This underscores a broader point: limitations often stem from specific designs, not the underlying AI.
Competitors like Figure AI, which unveiled their Figure 03 humanoid in October 2025 with advanced dexterous hands, could soon demonstrate orange-peeling or shirt-hanging with ease. Figure's model boasts improved manipulation, drawing from over 100 years of combined team expertise in AI and humanoids.
These developments come amid a robotics boom. PI raised $600 million in Series B funding in November 2025, valuing the company at $5.6 billion, backed by investors like CapitalG, Jeff Bezos, and OpenAI. Their π 0.6 incorporates "recap" techniques for better data efficiency, boosting success rates across tasks.
Broader industry trends show humanoid robots advancing rapidly: Tesla's Optimus Gen 2 handles factory tasks, while events like the Nebius Robotics & Physical AI Awards in December 2025 honored startups for foundation models and hardware innovations.
Looking to 2026, expect explosive growth. With robust hardware pipelines, efficient data cycles, and emerging reinforcement learning (RL) atop VLAs - PI skipped RL here, potentially capping success below 90% - companies will add skills quickly.
Pilot programs in warehouses and homes are already underway, signaling humanoids could become commonplace.
As Deloitte notes, AI-powered robots are evolving from rigid machines to adaptive ones, unlocking new frontiers in safety and complexity. T-shirt folding might soon feel quaint, replaced by robots mastering keys, sandwiches, and even pet cleanup. The Olympics aren't just games - they're a roadmap to a robotic revolution.
Also read:
- Hollywood's Horizon: Bold Predictions for 2026
- Bitmain's Drastic Price Cuts: Survival Tactics in a Struggling Bitcoin Mining Sector
- Navigating the AI Wave: What Professionals Really Think and Feel About AI in the Workplace
Thank you!
