Auryal/libero-microwave-grm-rollouts

LIBERO Microwave GRM Rollouts

Dense-reward-annotated rollout dataset from GRPO training of OpenVLA-OFT on LIBERO-10 Task 9 ("put the yellow and white mug in the microwave and close it").

Dataset

Stat	Value
Episodes	~2,100 (T >= 5 steps)
Format	LeRobot (parquet + images)
Task	put the yellow and white mug in the microwave and close it
Size	41 GB
Reward model	Robo-Dopamine GRM-3B
Policy	OpenVLA-OFT (LoRA, GRPO-trained)
Simulator	LIBERO (robosuite)
Robot	Franka Panda

Columns

Each parquet episode contains:

Column	Type	Description
image	uint8	Front camera (256x256x3)
wrist_image	uint8	Wrist camera (256x256x3)
state	float32	Proprioceptive state (8-dim)
actions	float32	Action (7-dim)
grm_score	float32	Per-step GRM progress score (0-1, cumulative mode)
grm_progress	float32	Accumulated GRM progress
env_sparse_reward	float32	Original sparse reward (1.0 on success)
is_success	bool	Episode success
dones	bool	Termination flag

Reward Visualizations

The dataset includes ~2,088 per-episode MP4 videos (reward_vis_*.mp4) showing:

• Top row: 3 camera panels (Front, Wrist L, Wrist R)
• Bottom row: Hop reward plot (symlog) + accumulated progress plot (linear)
• Task name and step counter overlaid

Training Context

This data was collected during a 20-epoch GRPO training run that exhibited reward hacking — success rate dropped from 86.5% to 0% while policy loss converged. The dataset captures the full trajectory of policy collapse, making it useful for studying:

• Dense reward model failure modes
• Reward hacking in embodied RL
• GRM scoring patterns across policy quality levels

Early episodes (epoch 1-5) show competent manipulation. Late episodes (epoch 15+) show degenerate behavior that still receives positive GRM scores.

Associated Model

Auryal/openvla-oft-libero-microwave-grm-grpo

Citation

@misc{auryal2026grm-grpo,
  title={Dense Reward RL with GRM for Robotic Manipulation},
  author={Auryal},
  year={2026},
}