Instructions to use ppprock11/so101_best_success with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- LeRobot
How to use ppprock11/so101_best_success with LeRobot:
# No code snippets available yet for this library. # To use this model, check the repository files and the library's documentation. # Want to help? PRs adding snippets are welcome at: # https://github.com/huggingface/huggingface.js
- Notebooks
- Google Colab
- Kaggle
best_success β ORIGINAL 35k checkpoint
Preserved baseline: exact weights, wrapper, configs and processors from the 35,000-step checkpoint. No inference.json override is added. Default temporal ensemble coefficient 0.01; one action per query; gripper bounds 5..45; snapping off.
Selected for highest observed success in our checkpoint comparison, not a claim of optimality or official qualification.
Local results (all retained): original three-seed test 29/30, pooled mean successful completion 6.81483s. Separate seed2026 repeat 10/10 at5.75s; seed2027 repeat 10/10 at7.94s. The repeats do not replace the original30 results. No complete repeat seed2028 yet. None establish the full acceptance criterion of >=28/30 with mean<5.5s for a complete30 test.
Training: stock LeRobot0.4.4 ACT;800 scripted simulation episodes; chunk30, batch8, lr1e-5, KL10. 35k total steps, resumed from25k with optimizer/RNG state. No human-demo fine-tune in this checkpoint.
Modified experiments are published separately at ppprock11/so101_best_success_modified_v1. Never interpret their measurements as this original baseline's results.
Attribution: ACT by Tony Z. Zhao, Vikash Kumar, Sergey Levine and Chelsea Finn (2023); Hugging Face LeRobot; NWaves SO-101 simulation task.
- Downloads last month
- 6