DexPolicy representative checkpoints

This repository contains a representative subset of the object-specific policies reported in the DexPolicy paper. The released pairs use the YCB mustard-bottle relocation task and isolate the paper's exploration-scheduling intervention under PPO, GRPO, and FPO.

Checkpoints

File	Method	Object	Seed	Exploration during training
models/ppo_mustard_baseline_seed0.zip	PPO baseline	YCB 006 mustard bottle	0	Learned policy standard deviation
models/ppo_mustard_dexpolicy_seed0.zip	PPO + DexPolicy	YCB 006 mustard bottle	0	Piecewise schedule from 0.20 to a final 0.04 plateau
models/grpo_mustard_baseline_seed0.zip	GRPO baseline	YCB 006 mustard bottle	0	Inherited, learned policy standard deviation
models/grpo_mustard_dexpolicy_seed0.zip	GRPO + DexPolicy	YCB 006 mustard bottle	0	Scheduled standard deviation, 0.10 -> 0.03
models/fpo_mustard_baseline_seed0.pt	FPO baseline	YCB 006 mustard bottle	0	Fixed standard deviation, 0.10
models/fpo_mustard_dexpolicy_seed0.pt	FPO + DexPolicy	YCB 006 mustard bottle	0	Linear schedule, 0.20 -> 0.05

The PPO and FPO pairs train for approximately 5M environment steps with matched settings within each pair. PPO uses an MLP actor mean; FPO uses a conditional-flow actor mean. The GRPO pair continues from matched object-and-seed checkpoints at 507,904 environment steps for another 250,000 steps. It uses the same reward, terminal success definition, curriculum stage, and return-to-go credit module; the exploration schedule is the controlled difference.

The PPO and GRPO .zip archives follow the Stable-Baselines3 checkpoint layout and include the policy, PyTorch variables, optimizer state, and library-version metadata. The FPO .pt files are PyTorch training checkpoints for the conditional-flow actor and critic. These policies are object-specific and are not presented as a single cross-object policy.

This is a compact representative release rather than the complete multi-object checkpoint suite.