Splitting a humanoid's upper and lower body into two controllers cuts upper-body joint error to about 0.1 radians on H1 and H1-2
Synopsis
Humanoid whole-body control usually drives the upper and lower body with one policy, so large arm motions can disturb balance and the two tasks interfere during training. JAEGER splits the body into two independent controllers that share observations and rewards, initializes the upper body with supervised learning, then trains the whole body with reinforcement learning while supporting both root-velocity and joint-angle command modes. In simulation comparisons on the H1 and H1-2 platforms, JAEGER's upper-body joint tracking error is lower than HumanPlus, ExBody and OmniH2O, with pose-mode joint errors about 30% to 60% of those methods.
Interpretation
Humanoid whole-body control is split into separate upper-body and lower-body controllers, each focused on its own task instead of one network carrying both balance and arm posture. Earlier single-controller designs learned upper and lower body in one policy, where upper-body tracking became overly conservative to accommodate lower-body stability and the two interfered during training. In root-based mode comparisons on H1 and H1-2, JAEGER's upper-body joint angle errors are 0.1093 and 0.0933 radians, below HumanPlus at 0.2187 and 0.2547 radians and ExBody at 0.1689 radians on H1.
The two controllers share observations and rewards, and their outputs are concatenated into target joint positions for joint PD controllers, keeping the body coordinated while training separately. Separate control needs a coordination mechanism, otherwise upper and lower body cannot complete a whole-body motion together. The problem is cast as a Dec-POMDP with two agents, where the lower-body policy uses a 15-step observation history and the upper-body policy a 3-step history, both receiving proprioceptive state, previous actions, root velocity targets and pose targets.
Training follows a staged curriculum: the lower-body base policy is trained alone, the upper body is initialized by supervised learning with KL regularization on the lower body, and reinforcement learning then continues at full pose amplitude. Continuing reinforcement learning on the whole body from the base policy made the upper body overly conservative and let an unconverged upper body pull the lower body off the base policy and lose balance. In ablations, JAEGER w.o. Init reaches 0.4088 radians upper-body joint error in H1-2 pose mode versus 0.1192 for full JAEGER, and the single-transformer JAEGER Single reaches 0.4693 radians upper-body error in H1-2 root-based mode.
Perspective
The reported scope is JAEGER's tracking-error comparison in IsaacGym simulation on the H1 and H1-2 platforms, plus real-robot demonstrations and lateral-force stability trials; lateral forces were roughly 70N to 210N, and the authors report balance or recovery in over 90% of trials.
The authors note that H1-2 shows larger tracking errors across methods because its upper body is heavier, and that pose mode has generally lower root velocities, so upper-body interference with lower-body balance is less pronounced than in root-based mode and the ablation gaps differ between the two modes.