Humanoid robotics gets attention for the most visible part of the machine: the legs, the gait, the ability to cross a room without falling over. But if you are running pilots, buying systems, or backing the companies that need to ship, that is the wrong proxy for capability.
The more important question is whether a robot can manipulate the world in front of it. Can it recognize an object it has not seen before, predict how it will slip or deform, handle contact without breaking the task, and keep working when the scene changes? In an interview published by Robotics & Automation News, Columbia professor and SceniX co-founder Yunzhu Li makes the case that this is the real bottleneck in robotics deployment. Locomotion may be the visible milestone, but manipulation is what determines whether a system is actually useful.
That distinction matters because it changes how you evaluate progress. A machine that can walk confidently across a floor may still fail at the job operators care about: picking, placing, sorting, packing, assembling, opening, closing, and recovering from messy real-world variation. In deployment terms, movement gets you to the workcell. Manipulation determines whether the robot can survive contact with the work.
From locomotion to manipulation: the pivot that changes the roadmap
Li’s framing pushes the conversation away from headline demos and toward the harder systems problem underneath them. The challenge is not simply that objects are different sizes or shapes. It is that the robot has to reason about materials, friction, contact dynamics, occlusion, and the way environments change as soon as the robot touches them.
That is a much tougher problem than straight-line navigation. A robot can often be tested on movement with relatively controlled success criteria. Manipulation is different: the task depends on the object, the surface, the lighting, the placement, and the interaction history. What looks like a simple pick-and-place demo can turn into a chain of failures once the environment stops matching the lab.
For operators, that means the question is not whether a system can execute a scripted motion. The question is whether it can handle variation. For investors, the implication is equally direct: the value is not in the most photogenic platform, but in the stack that can close the gap between a controlled demo and a reliable industrial workflow.
Why manipulation is harder than movement
The reason manipulation remains the gating factor is that the robot is no longer just moving through space. It is interacting with the physics of the environment.
That means the policy has to account for:
- unfamiliar objects and geometries
- material properties that alter grip and slippage
- contact events that are hard to model exactly
- scene changes caused by the robot’s own actions
- failure recovery when an item shifts, falls, or jams
Li’s emphasis on environmental understanding reflects a broader robotics reality: the world is not a static map. It is a changing system that reacts when touched. That is why a robot that appears competent in isolation can still struggle in a warehouse aisle, factory line, or lab bench where the edge cases come fast and the tolerance for error is low.
This is also why locomotion can become a misleading benchmark. A humanoid that walks well may create the impression of broad readiness, but a deployment team still has to ask whether it can grasp a slippery container, identify a shifted part, or adapt when the object does not sit exactly where the plan expected it to be. In physical AI, the hard part is not body control in the abstract. It is interaction under uncertainty.
Simulation as a force multiplier, not a side project
Li’s second major point is that simulation is central. That is not a nice-to-have slogan. It is a practical answer to the data and iteration problem in manipulation.
Real-world robotic data is expensive, slow to collect, and often dangerous to gather at scale. If the goal is to learn how a robot should behave during contact-rich tasks, then you need a way to test many scenarios quickly before sending hardware into the loop. Simulation gives teams a path to generate synthetic experience, test grasp strategies, stress policies against variation, and iterate on failure cases without burning through time or equipment.
For developers, this changes the shape of the product cycle. Instead of treating simulation as a prelude to real-world development, it becomes part of the core workflow:
- build and test manipulation policies in simulation
- expose the policy to varied contact and scene conditions
- identify failure modes before hardware trials
- transfer the best-performing behaviors into real systems
- tighten the loop with real data and repeat
The reason this matters operationally is speed and safety. If you can narrow the range of unknowns before deployment, pilots become more predictable. That does not eliminate the need for real-world testing. It does make the testing cheaper, faster, and more informative.
For investors, simulation is relevant for another reason: it creates defensible infrastructure. The companies that can build or integrate high-quality sim-to-real pipelines, realistic grasp modeling, and robust validation environments are not just selling a model. They are reducing deployment risk.
Deployment reality: what operators need now
This is where the conversation becomes practical.
Factories, warehouses, and labs do not buy aspirations. They buy throughput, uptime, safety, and maintainability. If a robotics stack cannot show reliable manipulation under production constraints, the system remains a demo regardless of how good it looks on video.
Operators should be asking a small set of hard questions before a pilot:
- How does the system handle object variation across shifts and sites?
- What happens when the scene changes in ways the model did not expect?
- How much of the performance has been validated in simulation versus on hardware?
- What safety checks exist before the robot enters a live workflow?
- Can engineers inspect, debug, and retrain the stack without rebuilding the whole system?
That last question is where toolchains matter. Deployment is no longer just a hardware procurement exercise. It is a developer-tools problem. Teams need instrumentation, dataset management, simulation environments, policy testing, telemetry, and workflow controls that let them understand why a robot failed and how to improve it.
The relevance of developer tools to robotics deployment has only grown as pilots move from experiments into operational settings. A stack that is hard to debug is a stack that is hard to trust. And if operators cannot validate behavior before scale-up, they will cap exposure no matter how ambitious the marketing.
Where to invest now
For investors and managers, Li’s framing points toward a narrower but more durable set of bets.
The first is manipulation-focused autonomy stacks. These are the systems built around grasping, contact-rich control, scene understanding, and recovery rather than around locomotion spectacle. If the market is still learning what kind of robotics will scale first, the safer assumption is that usefulness will come from task reliability, not from novelty in movement.
The second is sim-to-real infrastructure. That includes simulators, synthetic data generation, domain randomization, test harnesses, and evaluation layers that make deployment less guesswork and more engineering. In a market still shaped by pilot risk, anything that shortens the path from a lab policy to a field-ready workflow has value.
The third is tooling. Not the glamorous kind, but the unglamorous software that lets teams inspect failures, manage versions, tune policies, and certify behavior. If physical AI is going to scale into industrial environments, the winners will need the same kind of developer discipline that modern software teams already expect.
The practical filter is simple: does the company make manipulation more reliable, or just make the robot easier to show?
Milestones to watch over the next 12 to 24 months
The next phase of robotics will be measured less by broad promises and more by specific, testable gains. The most important milestones are likely to be incremental, but they will matter.
Watch for:
- better grasping robustness across unseen objects and cluttered scenes
- stronger sim-to-real transfer on contact-rich tasks
- faster iteration cycles from simulation to hardware validation
- clearer safety and recovery behavior when tasks fail mid-execution
- more transparent toolchains for debugging and deployment
If those metrics improve, the market will have a better signal that manipulation is moving from research challenge to operational capability. If they do not, then humanoid momentum may continue to outpace real deployment readiness.
That is why Li’s emphasis is useful now. It cuts through the excitement around moving bodies and returns the industry to the harder question: can robots do useful work in the physical world, repeatedly, under changing conditions?
For now, that answer will be determined less by how far a robot walks and more by how well it handles what it touches.
