Robotics’ next bottleneck isn’t walking — it’s manipulation, and simulation is becoming the platform

The latest wave of humanoid and factory-automation announcements has a familiar rhythm: bigger production targets, more capable demo videos, and a lot of confidence that general-purpose robots are finally on the cusp of scale. But an interview published this week with Columbia professor and SceniX co-founder Yunzhu Li points to a different constraint than the one most of the market is still optimizing around.

Li’s framing is not about making robots move better through space. It is about making them understand and handle the world once they get there. That means manipulation: recognizing objects, inferring material properties, predicting contact, and adjusting when the environment changes under force. In other words, the hard part is not whether a robot can cross a room. It is whether it can open a crumpled package, sort unfamiliar items, fold cloth, or place a tool without breaking it.

That distinction matters because it changes the product roadmap. If locomotion is the showcase, manipulation is the bottleneck. And if manipulation is the bottleneck, then simulation stops looking like a research convenience and starts looking like the core development platform.

Why now

The timing helps explain why Li’s comments land as more than just another academic take. Robotics coverage and investment chatter have intensified sharply in early June 2026, alongside a fresh round of promises about humanoids, warehouse systems, and industrial automation. The market is clearly re-rating robotics as an AI-adjacent category again.

But the center of gravity is shifting. The excitement is no longer only about whether a robot can stand, walk, or navigate. It is about whether developers can build systems that cope with the messy edge cases of physical interaction at scale. That is the point where a sim-first workflow becomes attractive: teams need more data, more coverage of rare events, and more controlled ways to test failure modes without breaking hardware or waiting for real-world logs to accumulate.

Li’s stance on simulation and manipulation fits that moment. It signals that the next phase of robotics tooling will not just be about better model training loops; it will be about creating development environments that let teams reason over the physics of interaction before they ever ship a robot into production.

What manipulation actually means

The word gets used loosely in robotics, but Li’s definition is much more demanding than “pick and place.” Manipulation forces a robot to reason about the identity of objects, the properties of materials, the dynamics of contact, and the way an action changes the scene.

That is a different problem from locomotion in at least three ways.

First, the robot has to infer what it is touching. A rigid box, a soft bag, a slippery bottle, and a deformable cable all require different policies even if they look similar from one camera angle.

Second, the robot has to predict how the world will respond. Push too hard and a stack collapses; pull at the wrong angle and a drawer jams; grip with the wrong force and the object slips or deforms.

Third, the robot has to adapt to change while acting. The object may move, the scene may shift, and the very interaction can alter the conditions that determine the next step.

This is why manipulation remains the stubborn wall for robotics commercialization. A robot can appear competent in a controlled environment and still fail in a warehouse aisle, a kitchen, or a factory line where objects vary, lighting changes, and physical interactions are less forgiving than navigation.

Simulation as the platform

Li’s emphasis on simulation is important because it reframes what simulation buys teams in practical terms.

At the most basic level, simulation lets developers generate data faster than they can collect it in the real world. That matters when the system needs exposure to many objects, surfaces, contact conditions, and environmental variations before it becomes useful. In a physical deployment, those scenarios are expensive, slow, and often unsafe to reproduce repeatedly.

Simulation also enables safer experimentation. Teams can probe failure modes that would be costly to discover on hardware: dropped objects, over-force contacts, unstable grasps, collisions, and edge cases involving unusual materials or cluttered spaces. For a product team, that can shorten iteration cycles and reduce the reliance on live robot fleets just to find basic weaknesses.

But simulation is not a free lunch. The obvious risk is the sim-to-real gap: what performs well in a virtual environment may degrade once real sensors, real friction, and real-world messiness enter the loop. That means the value of simulation depends on how seriously teams treat validation.

The most credible approach is not to pretend simulation replaces reality. It is to use simulation to cover the long tail of interaction cases, then close the loop with real-world testing that targets the failure patterns the simulator exposes.

What this means for tooling and product roadmaps

Li’s interview points toward a practical shift for robotics developers: build around manipulation first, and make simulation the organizing layer of the stack.

That suggests three priorities.

1. Treat simulation-to-deployment as an end-to-end pipeline.

Teams should not think of simulation as an isolated training environment. The useful stack connects scenario generation, policy training, evaluation, real-world logging, and retraining. The point is to move manipulation learning through a loop that keeps improving against realistic physical edge cases.

2. Define benchmarks around interaction, not just motion.

Navigation metrics are insufficient if the product sells task completion. Benchmarks need to capture grasp stability, contact robustness, material variability, recovery from failure, and how well a system handles scene changes after interaction begins. Otherwise teams risk optimizing for demos that look impressive but do not map to deployment risk.

3. Build data collection around rare failures.

The most valuable data in manipulation is often the least common data: slips, jams, deformations, partial occlusions, and unexpected object behavior. Simulation can help surface these cases systematically, but teams still need operational pipelines that capture them in the field and feed them back into model development.

For product leaders, the implication is clear. If your roadmap assumes robots will become useful primarily by getting better at locomotion or navigation, you may be aiming at the easier problem. The competitive advantage is more likely to come from systems that can reliably interact with the world once they arrive there.

What to watch next

The near-term signals to track are not just robot shipment announcements or headline partnerships. They are the infrastructure bets underneath them.

Watch for simulation platforms that position themselves as manipulation development environments rather than generic physics engines. Watch for open datasets and shared benchmarks that focus on contact-rich tasks, material diversity, and environment change. And watch for companies that can show a real loop between synthetic training and physical validation, especially in settings where failure is expensive.

That is where the market’s robotics bet is likely to be contested next. The companies that win may not be the ones with the flashiest walking demos. They may be the ones that make manipulation data-efficient, testable, and safe enough to ship.

Li’s interview is a reminder that robotics is not waiting on a small improvement in motion. It is waiting on a much harder leap in physical understanding. Simulation is increasingly the platform that could make that leap practical.