Video Friday: Spot Plays to Show Its Whole-Body Manipulation

We have spent the last decade obsessed with the brain of the machine. The industry has poured billions into LLMs and neural networks, often forgetting that a brain without a body is just a glorified calculator. But if you look at the latest developments coming out of labs like ETH Zurich and companies like Boston Dynamics, the physical form is finally starting to catch up to the intelligence. We are moving past the era of stiff, programmed movements and into the era of whole-body manipulation.

The Whole-Body Problem

In the robotics world, we talk a lot about "whole-body manipulation." For those of you building in the space, you know this is the holy grail. It’s the difference between a robot arm sitting on a pedestal and a humanoid that uses its entire weight, torso, and legs to lift a heavy tire. The RAI Institute recently showcased some progress here, but as I watched the footage, I found myself looking for the torso involvement. Human beings don't just use their hands; we lean, we pivot, and we use gravity.

Building a machine that understands its own center of mass while interacting with an external object is incredibly difficult. We saw this in recent testing where a robot was tasked with handling a tire. The vision system was still struggling with basic environmental noise. If a person walked past or a random object was on the floor, the depth image saw a "blob" and the robot lost its target. This is the reality check for builders: your AI can be brilliant, but if your vision stack can't distinguish a tire from a passerby, you don't have a product yet.

Sim-to-Real is Getting Better

The Robotic Systems Lab at ETH Zurich is doing some of the most honest work in the field right now. They are bridging motion imitation with reinforcement learning. The goal is simple but the execution is complex: use human motion data as a training baseline, but let the robot autonomously adapt when it hits terrain it has never seen before. For founders, this is the blueprint. You can't hard-code every possible scenario in a warehouse or a construction site. You have to build a system that knows how to learn on the fly without breaking its expensive hardware.

Speaking of hardware, I’ve been keeping an eye on the ANYbotics team. Their ANYmal X is being marketed as the "least likely to explode" quadruped. While that’s a bit of dark industry humor, it highlights a serious point for industrial builders. If you’re deploying in oil and gas or chemical plants, reliability isn't a feature—it’s the entire business model. A robot that works 90% of the time is a liability in these environments.

The Vision Stack and the Edge

I want to talk about the hardware layer for a second because I think people are sleeping on the commoditization of vision. Luxonis just dropped their OAK 4 FFC. It’s a tiny board, about 70mm by 53mm, that lets you run five cameras with onboard AI and stereo depth. The kicker? It calibrates from the scene, not a checkerboard. This is a massive win for drone and humanoid builders. It moves the compute to the edge, reducing latency and making the machine more responsive to its environment.

We saw the power of edge compute and efficient algorithms with the MonoRace team from TU Delft. They won the AI Grand Challenge by beating human FPV world champions using nothing but a single camera and an IMU. No massive server rack, no tethered data link—just an onboard neural network mapping pixels to motor commands. This is how we should be thinking about the future: lean, efficient, and fast.

Tactile Sensing and the Human Touch

One of my biggest gripes with the current state of service robotics is the lack of "feel." We have robots that can navigate a room, but they can't pick up a dirty dish without crushing it or dropping it. Robotiq, one of the veterans in the industrial space, is pushing a new tactile sensing gripper that I think sets the tone for the next year. We need machines that understand pressure and texture.

BruBotics is also working on compliant robot hands. Are they ready for prime time? No. But this is the kind of technology required for robots that will eventually work alongside humans. A rigid steel hand is a safety hazard; a compliant, soft-touch hand is a collaborator. If you are building in the humanoid space, stop worrying about the walking speed and start worrying about the grip precision.

The Reality of Deployment

We are seeing some interesting real-world deployments in China, specifically at the Yulong River Scenic Area. DEEPRobotics is running their Lite3 and LYNX M20 robots in tourism scenarios. It’s a good test bed, but it’s still very controlled. For the founders reading this: don't get distracted by the polished demo videos. The real money and the real impact are in solving the messy, unglamorous problems.

Take the monkeyBot 2.0 from bluZenit, for example. It’s designed to climb industrial piping. It’s not a sexy humanoid, and it won't win a dance-off, but it solves a specific, high-value problem in facilities maintenance. That is where the sustainable businesses are being built right now.

The Takeaway for Builders

If you're in the trenches building the next generation of robotics, here is what I’m seeing:

• Focus on Robustness: Stop building for the lab. If your vision system gets confused by a "blob" in a depth image, it's not ready for the real world.
• Edge AI is Mandatory: The days of tethered compute are over. If your robot can't process its environment locally and in real-time, it will never be safe.
• Tactile is the New Vision: We have enough cameras. We don't have enough sensors that can "feel" the difference between silk and steel.
• Bio-Inspiration Works: From the SAWbot’s energy-efficient swimming to ETH’s motion imitation, looking at how nature solves movement is still the most effective R&D strategy.

The hype around AI is great for funding, but the builders who win will be the ones who figure out how to put that AI into a body that doesn't fall over when the lights change or the floor gets a little dusty. We’re getting there, but let’s stay skeptical of the demos and keep our eyes on the telemetry.

Read the original at IEEE Robotics →