The robotics industry is having the wrong argument. Every week there is a new foundation model announcement, a new architecture breakthrough, a new benchmark result. The conversation is almost entirely about the AI. What model architecture. How many parameters. Which inference stack. Almost nobody talks about the data. This is a mistake. And a recent paper out of the University of Tsukuba makes the case more sharply than anything we have seen in the industry press. Researchers built a bilateral teleoperation system for low-cost robot arms and tested a straightforward question: does the type of data you collect during teleoperation change the quality of what the robot learns? The answer was not subtle. When imitation learning policies were trained without any force information, a pick-and-place task succeeded in 0 out of 25 placement trials. When force information was added to both the input and output of the training data, the same task succeeded in 25 out of 25. Zero to perfect. Same architecture. Same robot. Different data. The task in question involved picking blocks of varying widths and placing them accurately. Without force data, the policy could not reliably determine how firmly it was grasping an object or whether placement was succeeding. The robot was, in effect, flying blind through the physical dimension of the task. It could see. It could move. But it could not feel. This is not an edge case. It is the standard condition for most robot learning systems deployed today.
Position-only data is the industry default, and it is not enough
The dominant approach to teleoperation data collection right now is unilateral control. An operator moves a controller, the robot mirrors the motion, and the system records joint positions and camera images. This is simple to implement, inexpensive, and produces large volumes of data quickly. It also strips out half of what makes human manipulation competent. When a person picks up a fragile object, adjusts grip on a slipping tool, or presses a component into a tight socket, they are not just executing a position trajectory. They are continuously reading and responding to force. The fingertips are feeding back information about contact, compliance, and resistance. The motion adapts in real time to what the hands are feeling. Position-only data captures the motion. It does not capture the physics. A policy trained on position-only data learns to replicate motions. A policy trained on force-aware data learns to interact with the physical world. These are fundamentally different capabilities, and the gap becomes most visible exactly where industrial deployments are most demanding: contact-rich tasks, variable object properties, tight tolerances, recovery from unexpected states.
What this means for training data as a strategic asset
The foundation model companies building physical AI understand that data quality is the constraint. Physical Intelligence, Skild, and others are not limited by compute or architecture. They are limited by the availability of manipulation data that actually teaches robots how to handle physical uncertainty. Position trajectories are abundant. Force-aware, contact-rich demonstrations collected in real production environments are not. This is the gap that matters. And it is a gap that cannot be closed by collecting more of the same kind of data. Scaling position-only demonstrations does not produce force competence. It produces more confident motion replication, which fails the moment the physical world does not cooperate. At Telepath, every deployment generates bilateral force-feedback data from real production environments. Not lab benches. Not simulation. Actual industrial facilities, handling actual product variability, across actual shift cycles. The manipulation data produced by our operations is structurally different from what most of the industry is collecting, because the teleoperation infrastructure that generates it is different. The robotics industry will eventually figure out that the data question is the important one. The companies that solved it first will have an advantage that is very difficult to replicate from behind.
The practical implication for operators and AI teams
If you are evaluating a robot deployment partner, ask what their teleoperation infrastructure captures. Position and images is a starting point. Force, torque, contact state, and recovery behavior is what actually teaches a robot to work. If you are building physical AI and sourcing training data, ask whether the demonstrations were collected in real environments with real physical variability. A dataset of ten thousand clean lab demonstrations is not the same as a dataset of ten thousand production demonstrations, even if the task looks identical on paper. The model learns what the data contains. If the data does not contain physics, the model does not learn physics. That is the argument. It is not complicated. The industry just has not fully absorbed it yet.