knightnemo's picture
Upload README.md with huggingface_hub
3c3e812 verified
|
Raw
History Blame Contribute Delete
1.62 kB
metadata
license: mit
tags:
  - robotics
  - robomimic
  - diffusion-policy
  - dppo
  - reinforcement-learning

DPPO fine-tuned experts for robomimic (lift / can / square)

Teacher checkpoints used by the FlowDagger / DS-GRPO noise-space adaptation experiments in VAM_Learn_from_Human_Video (branch feat/efficient-ft).

Each checkpoint is a diffusion policy fine-tuned with the official irom-lab/dppo recipe (cfg/robomimic/finetune/<task>/ft_ppo_diffusion_mlp.yaml: DDPM 20 steps, ft_denoising_steps=10, state observations, seed 42) starting from the DPPO-released pre-trained (BC) checkpoints — DPPO does not release the fine-tuned models, so these were trained from scratch on CMU Babel.

task file arch (DiffusionMLP) ft iters success rate*
lift lift/expert.pt 512^3, time_dim 16 state_80 99.5%
can can/expert.pt 512^3, time_dim 16 state_150 100%
square square/expert.pt 1024^3 + cond_mlp[512,64], time_dim 32 state_200 97.5%

*200 stochastic episodes, sparse-success criterion, robomimic env stack from the DSRL fork (robosuite v1.4.1).

Format & usage

state_*.pt from DPPO fine-tuning: checkpoint["model"] holds actor.* (base) and actor_ft.* (fine-tuned head; use ft_denoising_steps=10 denoise routing). In the companion repo:

export DPPO_EXPERT_ROOT=/path/to/experts   # <task>/expert.pt layout
python -m algorithm.trainers.main algorithm=grpo_robomimic_can_expert