HiRE: Base Policies for Finetuning
Behavior-cloning checkpoints for HiRE policy finetuning on MimicGen and RoboMimic Tool Hang. Each task includes the selected checkpoint and a Hydra configuration, which the finetuning loader uses to restore the policy architecture. The three MimicGen configs are original; the Tool Hang config is reconstructed from its archived W&B training configuration (see Source).
| Task | Demonstrations | BC seed | Checkpoint |
|---|---|---|---|
| Stack Three | 200 | 42 | stack_three/checkpoint/state_75.pt |
| Threading | 30 | 42 | threading/checkpoint/state_100.pt |
| Three Piece Assembly | 200 | 42 | three_piece_assembly/checkpoint/state_75.pt |
| Tool Hang | 20 | 42 | tool_hang/checkpoint/state_500.pt |
The three MimicGen policies use diffusion with 96 × 96 images; Tool Hang uses flow matching with 240 × 240 images and 20 Euler integration steps. All four use ResNet-18 visual encoders, a 1D UNet action head, and an eight-step action horizon. The selected epochs are 75, 100, 75, and 500, respectively. The original checkpoint payloads are preserved, including both model and ema weights. The HiRE finetuning loader selects model.
Files
stack_three/
.hydra/config.yaml
checkpoint/state_75.pt
threading/
.hydra/config.yaml
checkpoint/state_100.pt
three_piece_assembly/
.hydra/config.yaml
checkpoint/state_75.pt
tool_hang/
.hydra/config.yaml
checkpoint/state_500.pt
manifest.json
Keep each .hydra/config.yaml with its checkpoint. manifest.json records source provenance, file paths, available training overrides, and SHA256 hashes.
Finetuning
From the HiRE code repository, after installing dependencies and preparing demonstrations:
source scripts/set_path.sh
python scripts/launch.py finetune stack_three \
--checkpoint /absolute/path/to/HiRE-release/stack_three/checkpoint/state_75.pt \
--seed 42
Select threading, three_piece_assembly, or tool_hang and the corresponding checkpoint path for the other tasks. To download and use the Tool Hang bundle through the HiRE launcher:
python scripts/download_checkpoints.py --tasks tool_hang
python scripts/launch.py finetune tool_hang --seed 42
These commands require a HiRE checkout with Tool Hang registered in configs/checkpoints.json. With an older checkout, download this bundle directly and pass --checkpoint /absolute/path/to/HiRE-release/tool_hang/checkpoint/state_500.pt. The code handles the Python class paths in the original Hydra configurations.
Processed demonstrations and their matching normalization.npz files are required separately. Use the same task, camera order, delta-action convention, and demonstration subset as the BC policy. Tool Hang uses 20 proficient-human demonstrations, camera order sideview_image, robot0_eye_in_hand_image, 240 × 240 images, 9-D proprioception, and 7-D delta actions. Follow the Tool Hang data guide to prepare matching data and normalization.
Source
The three MimicGen checkpoint files and configurations are copied without modification from JimmyHan2004/MimicGen-pretrain, at revision e0d681881f64771604353b5d2535c71d87e5d4a1.
The Tool Hang state_500.pt was supplied separately and is published without modification. Its Hydra configuration was reconstructed from the archived config of W&B pretraining run 72kac01d: top-level value fields were unwrapped and W&B logging metadata was omitted. It is not the original .hydra/config.yaml file. See manifest.json for the checkpoint and reconstructed-config SHA256 hashes.