LIBERO Latent-Risk Benchmark for Robot Learning
Robot learning research in Carnegie Mellon's Department of Electrical & Computer Engineering: training ACT policies in PyTorch and building a simulated benchmark of manipulation tasks with hidden safety risks.
- Trained ACT models, evaluating action chunk sizes, model architectures, and ResNet backbones
- Developed a 10-scene, 30-task LIBERO benchmark with parameterized MJCF and unsafe task combinations
- Resolved mesh, joint, and scaling issues; wrote Python verification scripts for penetration and simulation stability