jadechoghari/ckpt-policy-dpk

Evaluation Results

PushT Environment

Method	Avg Sum Reward	Avg Max Reward	Success %	Eval Time (s)	Eval Ep Time (s)
DOT	98.61	0.927	67.3%	949.14	0.949
Diffusion PushT	104.84	0.955	65.4%	730.87	1.462
vqbet_pusht	87.22	0.817	57.0%	465.85	0.932

PushT Keypoints Environment

Method	Avg Sum Reward	Avg Max Reward	Success %	Eval Time (s)	Eval Ep Time (s)
Diffusion	101.66	0.969	71.0%	63.35	0.127
DOT	129.63	0.921	44.7%	382.24	0.382

PushT Training

• Policy: DOT
• Dataset: lerobot/pusht
• Environment: PushT-v0
• Batch Size: 24
• Training Steps: 1,000,000
• Logging: Every 1,000 steps
• Evaluation Frequency: Every 10,000 steps
• Checkpoint Saving: Every 50,000 steps
• Random Seed: 100,000
• Workers: 24
• Mixed Precision (AMP): Enabled
• Device: CUDA

PushT Keypoints Training

• Policy: DOT
• Dataset: lerobot/pusht_keypoints
• Environment: PushT-v0
• Batch Size: 24
• Training Steps: 1,000,000
• Logging: Every 1,000 steps
• Evaluation Frequency: Every 10,000 steps
• Checkpoint Saving: Every 50,000 steps
• Random Seed: 100,000
• Workers: 24
• Mixed Precision (AMP): Enabled
• Device: CUDA
• Training Horizon: 30
• Inference Horizon: 30