Streaming Failure Forecasting

Robot failure monitors that keep learning after deployment

A robot policy that worked in the lab meets dimmer light, new objects, a bumped camera and worn grippers in the field. We study failure monitors that keep learning from how each episode ended, so they still raise the alarm when conditions drift.

00.510 s5 s10 sfrozen thresholdSFF thresholdSFFfrozenSFF alarm, 3.5 s
A logged simulation episode with gripper fingers that wore out unseen by the camera. Our monitor (SFF) has learned from earlier episodes and raises the alarm 3.5 s in; the same monitor kept frozen never does, and the episode fails at its time limit. Step through this episode.

Research

  1. On real robots, the failure monitor that keeps learning wins

    Tested on real robot arms run by four different AI models, SFF beat frozen monitors and detect-and-reset on all four when the robot's job changed without warning.

  2. First results: can a robot's imagination warn us before it fails?

    A robot policy that imagines the next moment of video passed our pre-registered test: the gap between what it imagined and what then happened flagged failing attempts. A much simpler check, whether the scene has stopped moving, came close on its own, and the two together did better than either.

  3. Next: can a robot's own foresight warn of its failures?

    Future work. A newer kind of robot policy predicts what its cameras will see next. We plan to test whether the gap between that prediction and what actually happens is an early warning of failure, and how it compares with the monitors we already have.

  4. From simulation to the SO-101

    Our real-arm rig is built and recording: a low-cost SO-101, three cameras at 30 fps and AprilTags that label every episode automatically.

  5. Failure monitors that keep learning after deployment

    A monitor that keeps learning from every episode outcome ranked failures better than a frozen one on 4 of 5 simulation test suites, including the pre-registered primary test.

  6. Early experiments: a language model driving two arms

    With no demonstrations and no training, Claude Opus 5.5 planned every move of two SO-101 arms and handed an eraser from one arm to the other.

  7. A better alarm is not always a better forecast

    How we keep our evaluation honest: ranking, false alarms and warning time, each measured on its own, so a timer can never pass as a forecaster.