HuB: Learning Extreme Humanoid Balance

• Human-motion tracking is commonly trained in simulation and then deployed on hardware, but this pipeline faces significant challenges on complex balance tasks.Reference poses are obtained from motion capture and retargeted to humanoid-specific motions before policy training.Source: §1 · p. 2
• Reference errors—including foot sliding from noisy pose capture and retargeting—make accurate motions essential for demanding balance tasks.HuB’s refinement pipeline uses designed initialization and post-processing to improve physical plausibility and transition stability.Source: §1 · p. 2
• Human–humanoid center-of-mass misalignment means strict reference tracking may not produce stable equilibrium, so HuB relaxes tracking and adds shaping rewards.These choices let the policy explore stable behavior near the reference while encouraging physically plausible motion.Source: §1 · p. 2
• Noisy IMU and VIO inputs and mismodeled ground contact can destabilize real-world balance, so HuB uses localized tracking, IMU-centric perturbations, and high-frequency pushes.These measures target VIO dependence, realistic sensor noise, and simulation discrepancies, respectively.Source: §1 · p. 2
• On the Unitree G1, HuB completes extreme balance tasks such as Swallow Balance and Bruce Lee’s Kick, while tracking-based baselines consistently fail.The baselines lose balance and fall or abandon single-leg motions; HuB also completes 10 consecutive executions in one rollout without intervention or resets.Source: §1 · p. 3

• Prior balance work focuses mainly on locomotion or transient stabilization, whereas HuB targets sustained quasi-static balance under extreme conditions.The target requires precise whole-body coordination and disturbance resilience.Source: §2 · p. 3
• Learning-based humanoid control spans locomotion and skills such as running, jumping, dancing, parkour, and loco-manipulation, but mostly emphasizes dynamic stabilization over precise quasi-static poses.HuB introduces a balance-centric framework for sustained stability in extreme configurations such as single-leg support.Source: §2 · p. 3
• Sim-to-real strategies include system identification, real-to-sim feedback, and domain randomization, but their effectiveness for humanoid quasi-static balance is underexplored.HuB builds on domain randomization with balance-specific perturbations.Source: §2 · p. 3

3.1 Background

• HuB’s tracking-based balance task is formulated as goal-conditioned reinforcement learning, with proprioceptive observations, reference-motion goals, desired joint-angle actions, and tracking-and-stability rewards.A low-level PD controller executes the policy’s desired joint angles.Source: §3.1 · pp. 3–4
• The pipeline extracts SMPL poses from video, retargets them to humanoid references, trains a privileged-observation teacher with PPO, distills an onboard-observation student with DAgger, and deploys the student on the robot.Policies are trained in simulation.Source: §3.1 · p. 4

3.2 Reference Motion Refinement

• Reference refinement uses SMPL-based joint initialization, grounded-foot correction, COM-based trajectory filtering, and extended double-support phases to improve motion accuracy, physical feasibility, and transitions.Trajectories are discarded when the ground-projected COM is more than 0.2 m from the support-foot center.Source: §3.2 · pp. 4–5

3.3 Balance-Aware Policy Learning

• Balance-aware learning relaxes reference tracking with σ = 0.6 m so policies can adjust COM when strict tracking would destabilize the humanoid.The tolerance permits fine-grained COM adjustments near the reference motion.Source: §3.3 · p. 5
• COM, foot-contact-mismatch, and close-feet shaping rewards encourage balance within the support polygon, intended foot contacts, and stable lower-body poses.The contact penalty discourages unintended ground contact by the non-supporting foot.Source: §3.3 · p. 5

3.4 Sim-to-Real Robustness Training

• For sim-to-real robustness, HuB uses local-frame tracking without odometry in training and deployment, avoiding VIO dependence and keeping the two settings consistent.The reference root is aligned with the humanoid’s current root pose.Source: §3.4 · pp. 5–6
• Student training perturbs root orientation with temporally correlated OU noise, while teacher training adds small high-frequency root-velocity pushes to model sensor noise and subtle instability.Pushes occur every 1 s at up to 0.5 m/s; the method contrasts them with prior low-frequency, high-magnitude pushes that are unsuitable for single-leg balance.Source: §3.4 · p. 6

4.1 Experiments Setup

• Experiments use the Unitree G1, IsaacGym simulation, 100 episodes per policy, and test-time root-velocity pushes and IMU noise; baselines share HuB’s motion data and keypoints.The baselines are adapted to localized tracking and trained from scratch.Source: §4.1 · pp. 6–7, §4 · p. 6

4.2 Simulation Results

• 100% success rate makes HuB the only method to complete the extreme simulation balance tasks; baselines almost always fail from unintended contacts by the non-supporting foot.HuB also has smaller slippage, shorter airborne durations, lower action variability, and lower tracking errors.Source: Table 1 · p. 7, §4.2 · p. 7
• A tracking tolerance of σ = 0.6 m balances tracking fidelity and stability; smaller values increase slippage, airborne time, and action rate, while excessively large values impair task success.The large-tolerance trade-off also includes higher tracking errors and more contact mismatches.Source: §4.2 · p. 7
• Removing the COM reward increases slippage and airborne time; removing contact-mismatch or close-feet penalties respectively raises incorrect contacts or reduces standing stability.Removing the contact-mismatch penalty also lowers success rates.Source: §4.2 · p. 7
• Global rather than localized training tracking, omitting IMU noise, or omitting high-frequency pushes degrades performance; low-frequency pushes particularly hurt tasks with narrow balance regions.The paper suggests infrequent large perturbations may be too hard for the humanoid to withstand.Source: §4.2 · pp. 7–8
• SMPL-initialized retargeting yields lower loss than non-SMPL initialization across tasks, with notably large reductions for Deep Squat.The comparison measures retargeting loss after 500 optimization steps.Source: Figure 3 · p. 8, §4.2 · p. 8

4.3 Real-World Results

• HuB completes real-world Swallow Balance and Bruce Lee’s Kick stably, while OmniH2O falls, shakes excessively, or abandons leg lifting in evaluated poses.Table 2 compares HuB and OmniH2O across real-world evaluation metrics.Source: §4.3 · p. 8, Table 2 · p. 8
• During a forceful soccer-ball strike, HuB recovers to a stable state with minimal corrective motion and completes 10 consecutive Bruce Lee’s Kicks without reset or intervention.The kick trials were performed in a single take.Source: §4.3 · p. 8

5 Conclusion

• HuB enables challenging balance poses that tracking-based baselines consistently fail to complete.The framework also demonstrates robustness to disturbances and consistency over long-horizon deployments.Source: §5 · p. 8
• HuB addresses reference-motion inaccuracies, balance-policy learning difficulties, and the sim-to-real gap in one framework.Source: §5 · p. 8
• HuB demonstrates robustness to disturbances and consistency over long-horizon deployments.Source: §5 · p. 8

6 Limitations

• Some HuB components rely on balance-task-specific assumptions and may not directly apply to jumping or parkour.Source: §6 · p. 9
• Adapting the trained policies to novel, substantially different motions typically necessitates retraining.Source: §6 · p. 9

A Real Robot Setup

• The real-world setup uses a 29-DoF Unitree G1, onboard IMU and joint encoders, and policy inference at 50 Hz.Inference runs on an onboard NVIDIA Jetson Orin NX.Source: A Real Robot Setup · p. 15

B.1 State Space Design

• The teacher policy has access to the full states required for reference tracking, while the student uses deployment-accessible observations and a 25-step history.The student uses 12 tracking keypoints at the hips, knees, ankles, shoulders, elbows, and wrists.Source: §B.1 · p. 15

B.2 Rewards

• The reward components are summarized in Table 5, whose hatted quantities denote reference variables and unmarked terms denote humanoid state variables.Source: Table 5 · p. 16, §B.2 · p. 16

B.3 Domain Randomization

• HuB’s domain-randomization strategies include high-frequency push disturbances intended to bridge the sim-to-real gap and improve balance robustness.Source: §B.3 · p. 17

B.4 IMU Noise

• Ornstein-Uhlenbeck noise is added to the IMU’s Euler-angle observations; its mean-reversion term limits excessive drift.The stated parameters are θ = 25 and σ = 250.Source: §B.4 · p. 17

B.5 Hyperparameters

• Table 7 presents the hyperparameters used for training HuB.Source: §B.5 · p. 17, Table 7 · p. 17

B.6 Foot Contact Labeling

• During grounded-foot correction, the foot with the lower ankle height is labeled as grounded.For the foot-contact mismatch penalty, the criteria in Table 5 determine which foot is grounded.Source: §B.6 · p. 18

C.1 Experiments Setup Details

• For a fair comparison, H2O and OmniH2O are trained from scratch using HuB’s balance-motion data and the same tracking keypoints.Source: §C.1 · p. 18
• Training and evaluation use the same domain randomization except for random external pushes.Source: §C.1 · p. 18
• External pushes reach 0.5 m/s during training and 0.1 m/s during evaluation.The passage describes the larger training pushes as promoting robustness to stronger disturbances and the smaller evaluation perturbations as closer to realistic deployment.Source: §C.1 · p. 18

C.2 Additional Results

• HuB outperforms baselines in completion, stability, and tracking errors across three additional tasks.The reported results are from simulation.Source: Table 8 · p. 18, §C.2 · p. 18
• The noise comparison distinguishes perturbing IMU orientation before computing policy observations from adding noise directly to those observations.The former is termed coupled; the latter independent.Source: §C.2 · p. 18
• Coupled OU noise achieves the best overall performance among the tested variants on Bruce Lee’s Kick.The comparison includes coupled Uniform noise and independent OU and Uniform noise.Source: §C.2 · p. 18, Table 9 · p. 19

C.3 Failure Analysis of Baselines

• Baseline failures primarily involve falling or foot-contact mismatch.Both-feet contact counts as failure because it means the single-leg task was not completed.Source: §C.3 · p. 19
• Swallow Balance failures involve falling, while Ne Zha Pose and Single-Leg Stand failures involve contact mismatch.These are the task-specific failure examples reported for the baselines.Source: §C.3 · p. 19
• Possible contributors to low baseline success include missing contact penalties, overly small tracking tolerances, and reward designs such as penalizing non-flat base orientations.The passage notes these are among many possible factors.Source: §C.3 · p. 19