# Reward Gating

Emotional video generation has no single correct target, and fixed fusion of visual, motion and text-alignment rewards captures emotional intent poorly. EmoReward distills reward traces from one completed training run into per-emotion-regime reliability estimates, converts them into bounded weight multipliers, and combines them with current-video instance gates in the next run without updating any reward-model parameters. On HunyuanVideo 1.5 with DanceGRPO, FlowGRPO and SAGE-GRPO, VideoAlign Overall rises under all three algorithms, including SAGE-GRPO with EmoReward moving from -0.8269 to -0.2624.

Emotional video generation has no single correct target, and fixed fusion of visual, motion and text-alignment rewards captures emotional intent poorly. EmoReward distills reward traces from one completed training run into per-emotion-regime reliability estimates, converts them into bounded weight multipliers, and combines them with current-video instance gates in the next run without updating any reward-model parameters. On HunyuanVideo 1.5 with DanceGRPO, FlowGRPO and SAGE-GRPO, VideoAlign Overall rises under all three algorithms, including SAGE-GRPO with EmoReward moving from -0.8269 to -0.2624.

This work presents a unified reinforcement-learning whole-body controller that takes a single 6-DoF end-effector target as the sole task-level command and directly outputs coordinated actions for a bipedal base and a six-joint arm across 14 joints, reaching 88.30% task success in simulation (2.85 cm mean position error, 5.38 cm P95), and on a real bipedal platform reusing the same controller across VR teleoperation, a learned diffusion policy, and scripted trajectories while extending vertical reach from roughly 38-163 cm for a floating-base plus inverse-kinematics baseline to roughly 3-191 cm.

This work presents a unified reinforcement-learning whole-body controller that takes a single 6-DoF end-effector target as the sole task-level command and directly outputs coordinated actions for a bipedal base and a six-joint arm across 14 joints, reaching 88.30% task success in simulation (2.85 cm mean position error, 5.38 cm P95), and on a real bipedal platform reusing the same controller across VR teleoperation, a learned diffusion policy, and scripted trajectories while extending vertical reach from roughly 38-163 cm for a floating-base plus inverse-kinematics baseline to roughly 3-191 cm.