Robotics Has Been Stuck for 70 Years

Deepak Pathak, Skild AI and Carnegie Mellon28:18 · Sept 2026 · 72K views
Thumbnail for Robotics Has Been Stuck for 70 Years Watch on YouTube
TL;DR
  1. 1

    Robotics has made less progress than other technology because it has been built around hardware and narrow pipelines instead of a general brain.

  2. 2

    Skild AI trains one omni-bodied model across robots and tasks by combining scalable pre-training data, high-quality teleoperation data, and deployment data.

  3. 3

    Robots still struggle with ordinary physical tasks such as stairs because they must understand changing surroundings, while isolated movements such as backflips are easier to simulate.

Summary

Deepak Pathak argues that robotics has been close to useful for decades because the field focused on building better hardware and hand-designed control systems. He compares modern demonstrations with robots from the 1950s and 1960s, then explains why robotics lacks the huge internet-scale dataset that drove language models. Skild AI's approach is to train one omni-bodied brain across humanoids, quadrupeds, arms, and other machines. The model uses scalable but imperfect data from simulation and human videos for pre-training, precise teleoperation data for post-training, and deployment data to improve future versions. Pathak shows examples including AirPods insertion with a gripper, omelet cooking on inexpensive arms using only a camera, stair climbing, GPU assembly, and adaptation after a robot loses legs. He is direct about the deployment problem: intelligence is only part of what makes robots difficult to use in factories or homes.

Key ideas
00:01

Robotics has repeated old demonstrations because it was treated as a hardware problem

Pathak opens by comparing current robotics excitement with demonstrations from decades ago. A robot that copies a block arrangement appeared in the 1960s, while a teleoperated leader-follower system shown in the talk dates to 1957. He says that many modern labs still collect robot data through teleoperation based on the same basic idea. Older systems could already perform tasks such as arranging objects, manipulating blocks, and playing games. His point is that robotics has not advanced like language, vision, phones, or chips because the field kept improving the body and the surrounding machinery without building a general brain.

04:27

Moravec's paradox explains why ordinary movement remains hard

Pathak uses Moravec's paradox to describe the mismatch between human difficulty and robot difficulty. Computers can beat people at mathematics and chess, but they still cannot reliably pick up chess pieces and arrange them on an arbitrary board. A backflip is easier for a robot because the robot only needs to control its own body in a known situation. Climbing stairs is harder because the robot must see the stairs, estimate their shape, and adjust to variation. Pathak says this makes robotics a foundational AI problem rather than a straightforward application of existing deep-learning methods.

05:42

Robot data cannot reach language-model scale through teleoperation alone

Language models benefited from enormous collections of internet data, but there is no equivalent internet of robot data. Pathak estimates that teleoperation takes about one minute to produce one example. Even if the entire US population collected data this way, reaching the scale of GPT-3 would take more than a century. Teleoperation is also expensive because it requires both a robot and a human, and it tends to produce many examples in the same setup. The field therefore needs data sources that can expand across tasks, environments, and robot bodies.

12:01

Robot data must balance scale, variety, and closeness to the machine

Pathak identifies three properties for judging robotics data: scalability, diversity, and closeness to the robot's joint angles. Simulation scales easily but requires people to build varied scenes. Human videos offer broad and inexpensive coverage but are far from robot actions. Teleoperation is close to the robot and high quality, but it is limited in amount and variety. He rejects the idea that one source is enough. The useful approach combines sources whose weaknesses differ, rather than looking for a single perfect dataset.

13:21

Skild's training recipe combines pre-training, post-training, and deployment data

Skild AI pre-trains its model on data that is scalable and diverse, including simulation and human videos, even when that data is lower quality or less directly connected to robot motion. It then post-trains with teleoperation data, which is more precise but scarce. A third source appears after deployment. Once robots operate in the world, their data can feed back into training. Pathak calls this a deployment data flywheel. He argues that one model must work across robot types for this loop to grow, because real deployments will use changing hardware rather than one fixed robot forever.

16:17

A general brain lets robots learn tasks from limited robot data

Pathak shows a gripper inserting an AirPod, a task that is difficult because the gripper cannot rotate the object at the end of the motion. The demonstration uses fake AirPods without magnets, so the robot must position them correctly. He also describes transferring skills from human videos with less than an hour of robot data. The robot can expand that experience through imagined scenarios inside its own model, which Pathak calls the robot's dream. This lets a general model use human demonstrations without requiring a large teleoperation dataset for every task.

18:56

Simple hardware can perform forceful tasks when the model reads the scene

The omelet demonstration uses arms costing about $4,000 and only a camera. There are no force sensors or other sensing hardware in the setup, even though cooking involves contact and force. Pathak says the system is fully end to end: it reads camera images and sends power to the motors, with only a proportional controller at the final frequency conversion. There is no hand-built state machine deciding when to start or stop. The robot continues when objects are changed, and Pathak says the behavior was trained with less than ten hours of data.

24:50

Deployment exposes problems that a lab demonstration does not

Skild's deployment examples include assembling GPUs for NVIDIA's Houston factory and carrying packages from a truck to a home's front door. Pathak stresses that real factory environments are noisy and disordered compared with carefully arranged demonstrations. A robot must also solve practical perception and common-sense problems, such as finding the front door. He says deployment is difficult for reasons beyond intelligence and hardware, but it must begin early if deployment data is going to improve the model.

26:25

An omni-bodied model can adapt when the robot itself changes

Pathak presents the shared brain as a safety advantage as well as a data strategy. If a robot loses a leg, its body has effectively changed, so the controller must adapt to a new configuration. In one demonstration, disabling legs lets the robot learn to walk on two legs in three trials. Another robot starts walking after a wheel is jammed. Pathak says the model can transfer to these bodies without being trained specifically on them. The intended behavior is adaptation to damage and unexpected changes rather than stopping because the original robot configuration no longer exists.

"When a robot is doing a back flip, it has to only know about its own body, nothing else."21:52
Who should watch
  • You are building robot-learning systems and need a practical way to combine simulation, human video, teleoperation, and deployment data.
  • Your team is deciding whether to build a controller for one robot or a model that can transfer across bodies and tasks.
  • You want to understand why a robot can perform a backflip yet fail at stairs, factory work, or other changing physical environments.