sled-umich/RACER-augmented_rlbench

---
license: mit
---

Rich language-guided failure recovery trajectories augmented from RLbench.

We gather the training and validation expert demos from RLbench as $D^{expert}$ (2250 episodes in total), perturb each episode five times and filter unsuccessful trajectories to obtain $D^{recovery+lang}$ (10,159 episodes in total). Both simple and rich language instructions are generated by prompting GPT-4-turbo for comparative study.

There are 18 tasks in total, 100 episodes for training set, 25 for validation set:
1. close_jar
2. meat_off_grill
3. place_shape_in_shape_sorter
4. put_groceries_in_cupboard
5. reach_and_drag
6. stack_cups
7. insert_onto_square_peg
8. open_drawer
9. place_wine_at_rack_location
10. put_item_in_drawer
11. slide_block_to_color_target
12. sweep_to_dustpan_of_size
13. light_bulb_in
14. place_cups
15. push_buttons
16. put_money_in_safe
17. stack_blocks
18. turn_tap

To run the model training, you need to preprocess this raw data into replay_buffer using [YARR](https://github.com/stepjam/YARR), or directly download we preprocess replay buffer from [here](https://huggingface.co/datasets/sled-umich/RACER-replay-public)