imishutk0410's picture
Add OGBench checkpoints (planners + invdyn for scene-play and cube-single)
b27b578 verified

OGBench Checkpoints

Scene-Play-v0 (Manipulation)

Best result: 48% GSC+Resampling+Pruning (50 episodes, seed=0)

Checkpoints

Component File Training Steps Notes
Planner (original) scene-play/planner/state_1495000.pt 1.5M energy_based_compdfu, batch=170, 8 GPUs
Planner (ogb_v1) scene-play/planner/ogb_v1_state_1495000.pt 1.5M Re-trained, batch=128, 4 GPUs. Better: 52% with same invdyn
InvDyn scene-play/invdyn/state_1600000.pt 1.6M invdyn_scene_h150, batch=32, horizon=150, uniform goal sampling

Eval Configs (50 episodes, seed=0)

Config Overall T1 (open) T2 (unlock) T3 (rearrange) T4 (drawer) T5 (hard)
GSC 36% 70% 40% 50% 10% 10%
GSC+Resampling (U=10,min=10) 40% 70% 20% 60% 50% 0%
GSC+Resamp+Pruning 48% 80% 50% 70% 30% 10%

Reproduction Commands

# GSC (baseline)
CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun --nproc_per_node=4 --master_port=29590 \
    -m src.compdiffuser.eval_sceneplay --env scene-play-v0 \
    --planner_name energy_based_compdfu --planner_epoch 1495000 \
    --invdyn_name invdyn_scene_h150 --invdyn_epoch 1600000 \
    --n_trials_per_task 10 --seed 0 \
    --ev_cp_infer_t_type gsc --ddim_steps 50 --cond_w 2.0 \
    --b_size_per_prob 40 --n_max_steps 1500

# GSC + Resampling (uniform U=10)
# Add: --ev_cp_infer_t_type gsc_resampling --num_resampling_steps 10 --min_resampling_steps 10

# GSC + Resampling + Pruning (best)
# Add: --ev_cp_infer_t_type gsc_resampling_pruning --num_resampling_steps 10 --min_resampling_steps 10 \
#      --pruning_start 0.5 --cv_threshold 0.01 --undo_eta 0.5 --use_gradient_ovlp --pruning_score_type inversion

Critical Training Notes

  • InvDyn batch_size=32 is essential. Batch=1024 gives 0-8%. The original invdyn_scene_h150 used batch=32.
  • InvDyn horizon=150 enables multi-step pick-place. Horizon=12 (default) gives 0%.
  • goal_sel_idxs must match plan_obs_select_dim: 12 13 14 19 20 21 26 27 28 29 32 33 36 38
  • Planner ogb_v1 (re-trained) gets 52% with same invdyn — better than original 46%.

Cube-Single-Play-v0

Best result: 28% GSC+Resampling (50 episodes, planner at 1.5M)

Component File Training Steps
Planner cube-single/planner/state_1495000.pt 1.5M
InvDyn cube-single/invdyn/state_1800000.pt 1.8M

Cube-single invdyn was trained with batch=1024 (needs retraining with batch=32 for better results).