haohw/dp_3wall_v5_mixed

DP 3Wall v5 Mixed Foundation

Best Diffusion Policy foundation checkpoint for the PushCube 3-wall / 3-pillar mixed-mode task.

• Local source path: /home/haohw/ManiSkill/examples/baselines/diffusion_policy/runs/dp_3wall_v5_mixed/checkpoints/best_eval_success_at_end.pt
• Uploaded file: best_eval_success_at_end.pt
• Size: 34,857,085 bytes
• SHA256: 18b17f9de85311884bc1224f8b0c5b5166281bdafc2fa43a87ecd0e4c0ec2e01
• Training data: 240 mixed demos across four modes (far_left, left_gap, right_gap, far_right)
• Architecture: state-based Diffusion Policy ConditionalUnet1D

Reference eval from local repro logs (/home/haohw/behavior-uncloning/logs/3pillar_dp_repro/baseline_mixed_seed0.log):

• 50 seeds, DDIM=50
• SR: 43/50 = 86.0%
• Mode dist: far_left=6 left_gap=31 right_gap=9 far_right=4 middle=0

Related handoff docs:

• /home/haohw/behavior-uncloning/docs/HANDOFF_DATA_AUDIT.md
• /home/haohw/behavior-uncloning/docs/3pillar_experiment_results.md