Video Friday: Digit Redecorates — 14 Real-World Clips, From a Couch Drag to a Grape Harvest Above 50 °C

Video Friday: Digit Redecorates

Source: IEEE Spectrum, Video Friday by Evan Ackerman, September 4, 2026

Video Friday is a weekly selection of robotics videos collected by the IEEE Spectrum robotics desk, along with a rolling calendar of upcoming events. This edition runs fourteen clips, and the throughline is unusually clear: almost every one of them is about a robot doing something in the real world rather than in a render. A humanoid carries a tray through the wobble of its own gait. A soft robot jumps forever under an infrared lamp. A quadruped hauls grapes through ground heat above 50 °C. Students ship their first sim-to-real behaviors in a month on an open humanoid platform.

It opens with Agility's Digit dragging a couch across a living room, which is the kind of household redecorating nobody asks a robot to do and everybody will eventually want one to do.

Videos out of Agility can be a little silly. This one is not: the couch drag is genuinely impressive, because it is a heavy, awkward, ungraspable object being moved by a biped that has to keep its balance while pushing through friction that changes with every centimetre of carpet.

Upcoming robotics events

Event	Dates	Location
Humanoids Summit Seoul	22–23 September 2026	Seoul
IROS 2026	27 September – 1 October 2026	Pittsburgh
CoRL 2026	9–12 November 2026	Austin

SteadyTray: keeping a payload level on a walking humanoid

Carrying things is the bottleneck nobody puts on a demo reel. A humanoid in dynamic locomotion is permanently falling and catching itself, and everything it holds inherits that oscillation. SteadyTray's ReST-RL attack on the problem is architectural rather than brute-force: decouple locomotion from payload stabilization in a hierarchical reinforcement-learning stack, and let each layer do one job.

Stabilizing unsecured payloads against the inherent oscillations of dynamic bipedal locomotion remains a critical engineering bottleneck for humanoids in unstructured environments. To solve this, we introduce ReST-RL, a hierarchical reinforcement-learning architecture that explicitly decouples locomotion from payload stabilization. Successfully deployed on the Unitree G1 humanoid hardware, this modular approach demonstrates highly reliable zero-shot sim-to-real generalization across various objects and external force disturbances.

The claim that matters is zero-shot sim-to-real across different objects and external force disturbances. Unsecured payloads are the realistic case in a warehouse or a hospital corridor; a strapped-down box is a much easier problem, and a lot of humanoid manipulation demos quietly assume one.

Figure: scaling compute, and what that is not supposed to solve

Figure's clip is about scaling compute so that its robots can have their compute scaled. The interesting part is the line that sits underneath it, which cuts against the prevailing framing of home robots as a data problem:

Solving for a robot in every home is not a data-and-compute problem, it's a safety-and-cost problem.

That is a substantive position from a company building exactly the stack the sentence says is insufficient on its own. More compute buys capability; it does not buy the failure-mode budget a household will tolerate, nor the unit price at which a family buys a second one.

RAI Institute: a handheld gripper with two thumbs per hand

The single most important thing to know about this gripper is that koalas have two thumbs on each hand. RAI's handheld data-collection device borrows that arrangement, and the point of the whole exercise is cheap, high-quality manipulation data gathered by people rather than by teleoperating a robot.

Opposed-thumb geometry matters for the data, not for the novelty: grasps collected with a device that is kinematically close to the target hand transfer far better than grasps collected with a parallel-jaw proxy.

DARPA Triage Challenge: finals in November

The DARPA Triage Challenge runs to its finals in November. The premise is prehospital emergency care under autonomy: robots and robotic teams triaging casualties, which means perception through blood, occlusion, and debris, plus decisions that have to be legible to a medic.

CMU APEX: adaptive full-body maneuvers instead of choreography

Most impressive humanoid footage online is carefully choreographed. Carnegie Mellon's Safe AI Lab is working the other side of that gap: teaching a humanoid to adapt as it moves, so an obstacle becomes something it routes around with its whole body rather than something that stops the take.

Online, humanoid robots are very impressive to watch, but behind the scenes, most of those movements are carefully choreographed. Researchers in Carnegie Mellon University's Safe AI Lab are instead teaching robots how to adapt. Their system, called APEX, allows a humanoid robot to navigate obstacles using adaptive, full-body maneuvers.

Adaptation is also the honest framing for safety. A robot that only ever executes a rehearsed trajectory is safe in the rehearsal and unknown everywhere else; a robot that can re-plan its whole body mid-step has a mechanism to deal with the case that actually hurts people.

NC State: teardrop soft robots that jump as long as the light is on

North Carolina State University built teardrop-shaped soft robots that leap upward or forward under infrared light and keep jumping for as long as the light is present. No controller, no battery in the body, no onboard sensing: the light is both the power source and the clock.

The robots are made of a liquid-crystal elastomer ribbon shaped like a teardrop, with a thin aluminum tube shaped like a V at one end. When exposed to light from an infrared lamp, the surface of the ribbon contracts, causing the ribbon to rotate. The stiff V at one end of the robot prevents the ribbon from simply rolling in place, causing the ribbon to twist tighter and tighter. This stores energy until the twist reaches a critical point when the ribbon releases that energy, causing the V at one end of the teardrop to snap downward and strike the surface. This launches the teardrop into the air.

The mechanism is a relaxation oscillator built from material properties rather than electronics: contract, twist, store, hit the critical point, snap, repeat. It is the cleanest possible demonstration that a gait can live in the body instead of the controller.

DARPA Lift Challenge: heavy-lift drones that surprised the desk

The Lift Challenge recap comes with an admission: expectations were low, and the heavy-lift drone designs turned out to be creative enough that a 2028 return is now something to look forward to.

Heavy lift is where airframe choices stop being incremental. Payload fractions that move the needle come from distributed propulsion, unusual structural layouts, and ground-effect tricks, not from a slightly better motor, which is why a competition format surfaces ideas that procurement would not.

Christian Hubicki: talking sense into the internet

Christian Hubicki, who coaches the Georgia Tech humanoid team, made a video attempting to talk some sense into the internet about what these robots can and cannot do. It is nearly three minutes of a researcher correcting the record, and it lands better coming from someone who is on the field with the hardware.

University of Tokyo: 44 sensing elements in a joint-covering skin

Proprioception in humans is not only muscle signals; stretched skin around a joint is a cue too. Tokyo's musculoskeletal humanoid Musashi-W now has a three-layer joint-covering skin carrying 44 pressure- and stretch-sensitive elements to mimic exactly that.

Humans use not only muscle signals but also stretched skin around joints as a cue for proprioception. To mimic this biological mechanism, we developed a three-layer joint-covering skin with 44 pressure- and stretch-sensitive elements for the musculoskeletal humanoid Musashi-W.

It is not a replicant, but one day, it will be. The engineering argument underneath the line is real: musculoskeletal humanoids have tendon-driven compliance that joint encoders only partly describe, and skin over the joint recovers contact and stretch state that the drive train cannot see.

LimX Dynamics: solve a mobility problem with an industrial arm

Having mobility issues with your robot? Staple it to the end of an industrial robotic arm. Problem solved. LimX put its humanoid on a fixed-base arm so the arm handles the coarse workspace and the humanoid handles the fine manipulation inside it.

Under the joke is a legitimate architecture question. Locomotion is expensive in energy, reliability, and safety certification; for a fixed workstation with a large reachable envelope, a mobile base may be the wrong purchase. Composition of two mature subsystems can beat one immature integrated machine.

Sharpa: the manager question

Sharpa's clip demos a dexterous hand system at work, and the framing question is the best one in the set: but what if I am the sort of person who needs to speak to a manager?

DEEP Robotics: a quadruped in the City of Fire

In Turpan, China, known as the City of Fire, summer ground temperatures can exceed 50 °C. During the grape harvest, farmers traditionally carry heavy baskets back and forth in that heat, and every extra minute in the sun costs fruit freshness. This year the DEEP Robotics Lynx M20S joined the harvest.

In Turpan, China — known as the City of Fire — summer ground temperatures can exceed 50 °C. During the grape harvest, farmers traditionally carry heavy baskets back and forth under the intense heat, while every extra minute in the sun can affect the freshness of the fruit. This year, the DEEP Robotics Lynx M20S joined the harvest.

This is the strongest deployment story in the edition, because the value is measurable in both currencies at once: human heat exposure goes down and shelf life goes up. Agricultural logistics is also terrain a legged platform handles better than a wheeled one, between vineyard rows, irrigation channels, and soft soil.

ROBOTIS: a student cohort ships sim-to-real in one month

The OH! GYM! Project's first cohort of university and graduate students developed and deployed their own humanoid behaviors on the open-source AI Sapiens K1 platform, going from motions created in simulation to a physical robot through repeated sim-to-real experiments inside one month.

This video showcases the achievements of the first OH! GYM! Project cohort, a group of university and graduate students who explored, developed, and deployed their own humanoid behaviors using the open-source AI Sapiens K1 platform. Over the course of one month, the students experienced the complete process of humanoid development — from creating motions in simulation to transferring them onto a physical robot through repeated Sim2Real experiments.

One month from simulation to hardware is a statement about tooling more than about talent. When the simulator, the policy stack, and the robot come from one documented open platform, the loop that used to eat a semester compresses to weeks, which is how a field grows a workforce.

What this edition adds up to

Read together, the fourteen clips separate into three groups. The first is locomotion meeting the physical world: Digit and a couch, SteadyTray and a tray, APEX and an obstacle, LimX and an industrial arm. The second is sensing and embodiment research that changes what the robot can perceive about itself, from NC State's light-driven soft jumpers to Musashi-W's 44-element joint skin. The third is deployment and pipeline: Turpan grapes, the DARPA challenges, Hubicki correcting expectations, and a student cohort shipping sim-to-real on an open platform.

The pattern worth noting is where the impressive footage sits. The clips with real operating constraints in them — heat, payload, unstructured terrain, unscripted obstacles — are the ones that are hardest to produce and the easiest to believe. Choreography is cheap. Turpan is not.

References

• Original column: Video Friday: Digit Redecorates, IEEE Spectrum, Evan Ackerman
• Agility Robotics · SteadyTray · Figure · RAI Institute
• DARPA Triage Challenge · DARPA Lift Challenge · CMU · NC State
• Christian Hubicki · University of Tokyo (Musashi-W) · LimX Dynamics · Sharpa
• DEEP Robotics · ROBOTIS OH! GYM!

Adapted for RobotWorld from IEEE Spectrum's Video Friday. All videos are the property of their respective sources and are hosted here for editorial commentary.