Robots Learn from Good Data, but Humans Benefit from Terrible Data Too

Summary

Chris Paxton discusses how robots primarily learn from good data, while humans benefit from learning from poor data, enabling faster task adaptation. The post introduces oopsie-data, an open dataset of failure trajectories, which is crucial for research in world models and reward function learning in robotics.

Source: Chris Paxton — X Robotics · Read original article ↗

Article text · Original source · English

Robots (and ai, more broadly) mostly learn from good data. Perhaps the biggest advantage humans have is that we learn from terrible data too -- it means we scale so much faster to new tasks.

Implications for data suppliers

RoboSignal interpretation and collection questions, not statements of buyer demand.

Confirm the required data type and collection setting with the buyer; this source does not establish a complete collection specification.
Validate demand and acceptance criteria with a buyer before scaling. Publication, popularity and a research result do not establish a purchase commitment.

Quoted postChris Paxton@chris_j_paxton

A lot of robotics data looks like this -- it's the result of a policy rollout that just doesn't accomplish anything. Right now, much of this data is just thrown out, deleted - it's gone forever. And yet for research in world models and reward function learning -- some of the most promising areas for the future of reliable, real-world robotics -- we need it. Enter oopsie-data, by @c_voelcker, @ArpitBahety, and @zhiyuan_zhou_: an effort to build the first open dataset containing failure trajectories. Learn more on @RoboPapers ->

View the quoted post on X