| --- |
| license: bsd-3-clause |
| tags: |
| - isaac-lab |
| - reinforcement-learning |
| - robotics |
| - peg-in-hole |
| - contact-rich-manipulation |
| - tacinsert |
| --- |
| |
| # TacInsert Checkpoints |
|
|
| This directory is reserved for optional pretrained checkpoints used for quick simulation smoke tests. Checkpoint files are large binary artifacts and are intentionally ignored by git. The released checkpoints are hosted on Hugging Face: |
|
|
| ```text |
| https://huggingface.co/tOMRmA/TacInsert-checkpoints |
| ``` |
|
|
| Download them into this directory before running the smoke-test commands below. |
|
|
| ```powershell |
| hf download tOMRmA/TacInsert-checkpoints TacInsert-LHole-III-Direct-v0.pth ` |
| --repo-type model ` |
| --local-dir source/TacInsert/TacInsert/tasks/direct/tacinsert/checkpoints |
| |
| hf download tOMRmA/TacInsert-checkpoints TacInsert-Manipulation-Square-SingleHole-Direct-v0.pth ` |
| --repo-type model ` |
| --local-dir source/TacInsert/TacInsert/tasks/direct/tacinsert/checkpoints |
| ``` |
|
|
| ## Released Checkpoints |
|
|
| The following checkpoints were used to validate the current TacInsert simulation code: |
|
|
| | File | Task | Observation dim | Notes | SHA256 | |
| | --- | --- | ---: | --- | --- | |
| | `TacInsert-LHole-III-Direct-v0.pth` | `TacInsert-LHole-III-Direct-v0` | 23 | L-hole Tol. III with contact-force observation and `fixed_sigma: True` | `FC158B0D72AFDAB0D860CEB379F8E3EC2B0B693D49C9B24629608A5BEA4B58AC` | |
| | `TacInsert-Manipulation-Square-SingleHole-Direct-v0.pth` | `TacInsert-Manipulation-Square-SingleHole-Direct-v0` | 26 | ManipulationNet-style square/rectangle single-hole sampler policy with contact-force observation, tolerance one-hot, and `fixed_sigma: False` | `0301CD4C57848EDA73FAE2D9F019D879C6D8AA1F29315CCA8A12BDC5B36CA10E` | |
|
|
| ## Smoke-Test Commands |
|
|
| Run from the repository root after installing the TacInsert extension. |
|
|
| ### L-Hole Tol. III |
|
|
| ```powershell |
| python scripts/rl_games/play.py ` |
| --task TacInsert-LHole-III-Direct-v0 ` |
| --num_envs 128 ` |
| --headless ` |
| --checkpoint source/TacInsert/TacInsert/tasks/direct/tacinsert/checkpoints/TacInsert-LHole-III-Direct-v0.pth |
| ``` |
|
|
| Expected behavior: the terminal prints one summary per completed episode. In local validation with 128 environments, three episodes reached approximately 94-98 percent success. |
|
|
| ### ManipulationNet-Style Square Single-Hole |
|
|
| ```powershell |
| python scripts/rl_games/play.py ` |
| --task TacInsert-Manipulation-Square-SingleHole-Direct-v0 ` |
| --num_envs 128 ` |
| --headless ` |
| --checkpoint source/TacInsert/TacInsert/tasks/direct/tacinsert/checkpoints/TacInsert-Manipulation-Square-SingleHole-Direct-v0.pth |
| ``` |
|
|
| Expected behavior with the default fixed Tol. I sampler: the terminal prints one summary per completed episode. In local validation with 128 environments, three episodes reached approximately 98-99 percent success. |
|
|
| To evaluate another fixed tolerance, edit `multi_hole_sample_weights` in `TacInsertManipulationSquareSingleHole` before running playback. For example, Tol. IV evaluation uses: |
|
|
| ```python |
| multi_hole_sample_weights = [0.0, 0.0, 0.0, 1.0] |
| ``` |
|
|
| In local Tol. IV validation with 128 environments, three episodes reached approximately 61-63 percent success. |
|
|
| ## Contact-Force Logging Check |
|
|
| For tasks with `contact_force["log_contact_force"] = True`, CSV logging is enabled only in evaluation mode with `--num_envs 1`: |
|
|
| ```powershell |
| python scripts/rl_games/play.py ` |
| --task TacInsert-Manipulation-Square-SingleHole-Direct-v0 ` |
| --num_envs 1 ` |
| --headless ` |
| --checkpoint source/TacInsert/TacInsert/tasks/direct/tacinsert/checkpoints/TacInsert-Manipulation-Square-SingleHole-Direct-v0.pth |
| ``` |
|
|
| CSV files are written to `contact_force_logs/`, which is ignored by git. |
|
|
