Robotics data startup @axisrobotics, which counts Pi Network Ventures from @PiCoreTeam among its seed investors, has published TUCO, a method for choosing which simulated demonstrations to use when training a robot. The research suggests that selecting the right data can matter more than collecting more of it.

TUCO stands for Trajectory-level Utility and set-level Coverage Optimization. The authors describe it as the first systematic study of data curation for sim-to-real robot policy co-training. The paper lists authors from Stanford University, Nanyang Technological University, AXIS Robotics, Carnegie Mellon University, and the University of California, Berkeley, among others, according to the project page .

86% Success With 600 Curated Demos

In the real-robot tests, each task was learned from only 10 teleoperated real demonstrations, which on their own produced a 16.7% average success rate. The researchers then added simulated demonstrations of the same task. They used either all 1,200 generated in simulation, or 600 chosen by a curation method, at the same real-to-simulation mixing ratio.

The gap between approaches was wide. Adding all 1,200 simulated demos lifted average success from 16.7% to 25.8%, a random half gave the same 25.8%, and the half selected by TUCO gave 85.8%. The setup covered three tasks, with 40 trials per policy . TUCO's per-task scores were 82.5, 75.0 and 100.0, for a macro-average of 85.8%.

The authors add a caveat on scope. The study uses a prescribed subset budget and source-to-target mix, so it does not determine how many source examples to retain. The results also come from a small number of tasks, and the work is a preprint on arXiv.

Beyond the real-robot tests, the paper reports simulation-based results too. TUCO matches full-data training on the Lift task with only 25% of the original training set and surpasses it on Square with 50%.

Axis Robotics and Pi Network Ventures

Axis Robotics builds data infrastructure for robot training. The company raised $12 million in seed funding led by Hack VC, with participation from Nomad Capital, Pi Network Ventures, 10K Ventures and several angel investors, and announced the round on July 27. According to TokenPost, Axis did not disclose Pi Network Ventures' individual contribution, and the fund is a $100 million initiative established through the Pi Foundation and launched in May 2025.

For the wider Pi Network ecosystem, the research is a sign of what a portfolio company is working on. The paper itself does not describe any use of the Pi blockchain or the $PI token.

Sources

TUCO: Curating Simulation Demonstrations for Sim-to-Real Robot Policy Co-Training (arXiv)

TUCO project page

TokenPost: Axis Robotics and Pi Network integration plan