tfayiz/arc-ai-sota-benchmarks
收藏资源简介:
--- task_categories: - robotics tags: - manipulation - mujoco - franka - expert-demonstrations - adversarial-testing size_categories: - 1M<n<10M license: apache-2.0 --- # ARC-AI SOTA Benchmarks Complete benchmark results from 19.74 GPU-hours of NVIDIA A100 simulation. ## Contents - `results/sota_full_results.json` — Raw data (all phases A-I) - `REPORT.md` — Production-grade validation report - `STRESS_TEST_REPORT.md` — Infrastructure stress test (76 min) ## Key Results | Metric | Value | |--------|-------| | GPU Throughput | 8.97M samples/sec | | Parallel Envs | 131,072 | | Adversarial Scenarios | 24 | | Physics Steps | 10M (zero failures) | | Total GPU Compute | 19.74 hours |
The ARC-AI SOTA Benchmarks dataset is a robotics dataset focused on manipulation tasks, utilizing the MuJoCo simulation environment and Franka robot platform. It includes expert demonstrations and adversarial testing scenarios, with a dataset size ranging from 1 million to 10 million samples. The dataset provides complete benchmark results from 19.74 GPU-hours of NVIDIA A100 simulation, encompassing raw data, validation reports, and stress test reports. Key performance metrics include high GPU throughput, large-scale parallel environment simulations, multiple adversarial scenarios, and zero-failure physics steps. This dataset is designed to support research and evaluation in robot learning and is licensed under Apache-2.0.




