Reinforcement learning (RL) performs strongly on continuous control benchmarks, yet many real-world systems such as articulated robots, musculoskeletal models, and dexterous hands require controlling tens to hundreds of actuators. In this regime, action dimensionality can be a dominant bottleneck, but existing benchmarks often entangle dimensionality with changes in objectives, morphology, or contact dynamics, obscuring how algorithms truly scale.
We introduce HiGym, a high-dimensional continuous control benchmark built on DMControl. HiGym comprises six environment families with multiple tasks and six embodiment complexities. We isolate action-space scaling by keeping tasks and dynamics as similar as possible while increasing degrees of freedom via a principled embodiment-scaling procedure that recursively splits joints to reach target dimensionalities. This yields matched task variants that differ primarily in action dimension.
Across PPO, TD-MPC2, and TD-M(PC)2, we find that model-based methods degrade substantially less as dimensionality increases, preserving higher sample-efficiency and asymptotic performance than model-free baselines. Among model-based methods, we find that TD-M(PC)2's policy-constraint loss provides a consistent further gain at high DoF locomotion tasks by suppressing value overestimation that arises from a structural mismatch between the learned policy and planner's behavior policy.
