THE ESSENTIALS IN 30 SECONDS

  • Demos included unpacking a 10 kg box of ball bearings, correcting a grasp that picked up paper with a bearing, and moving automotive radar units after their position was changed.
  • Rhoda's Direct Video-Action approach predicts short future video clips from recent observations and converts them into robot actions with a separate inverse dynamics model.
  • CEO Jagdeep Singh says a new task needs about 10–20 hours of teleoperation data; Humanoids Daily notes the visit does not prove sustained factory reliability.

Robot demonstrations usually show a task going right. A new video from Rhoda AI concentrates instead on what happens when things go slightly wrong. In an episode of the series "Field Trips with Schrep", host Mike Schroepfer visited the company and its founder and chief executive, Jagdeep Singh, as reported by Humanoids Daily on November 1.

What the robots did

The main demonstration involved unpacking a 10 kg box of ball bearings. A pair of arms opened the packaging, separated paper from a plastic liner, poured the bearings into a metal tote and set aside the waste. When the robot grabbed a bearing together with a piece of paper, it put the bearing back and picked up the paper again. Singh and Schroepfer also dropped crumpled paper into the tote to see whether the robot would find and remove it, and the robot used both arms to turn over a bag so it could pull out trapped paper.

A second task moved automotive radar units from a conveyor to an inspection fixture and then to an output slot. When Singh shifted a unit or turned it, the robot adjusted its grasp. A separate segment on electronics component reels showed teleoperated data collection rather than autonomous work.

How the system works

Rhoda calls its approach Direct Video-Action. According to the company's research page, a video model trained from scratch on web footage predicts a short clip of what should happen next, based on a long history of camera frames, and a separate inverse dynamics model turns that predicted clip into motor commands. The loop repeats several times per second. Singh told Schroepfer that adapting to a new task takes about 10 to 20 hours of teleoperation data on top of pre-training. Model inference currently runs in the cloud, with low-level control on the robot, and the models shown were not yet compressed for speed.

Limits and next steps

The demo robots use two seven-joint arms with two-finger grippers. For production, Singh describes wheeled mobile bases, which he says suit industrial customers worried about complexity and safe stopping. Rhoda aims for paid production deployments before the end of 2026, backed by a remote operations center. Humanoids Daily cautions that recovering within a filmed task is not the same as running reliably on a factory floor for weeks, which remains to be shown.