Humanoid Dodgeball: Learning Fast Dynamic Obstacle Avoidance Motions for Humanoids
Authors: Aoqian Zhang, Chunzheng Wang, Haoran Chen, Annan Tang, Cheng Xiang, Fan Shi, Shuzhi Sam Ge
Organizations: Department of Electrical and Computer Engineering, National University of Singapore, Singapore. · Shenzhen EngineAI Robotics Technology Co., Ltd., China. · School of Information and Communication Engineering, Dalian University of Technology, China.
Humanoids have become increasingly agile and capable in real-world environments. To operate safely in dynamic scenarios, they must rapidly evade fast-moving obstacles while maintaining stability and physically feasible whole-body motion. Learning such behaviors remains challenging due to the high-dimensional dynamics of humanoid systems and the need to jointly balance reactive obstacle avoidance, motion naturalness, and control robustness. In this work, we present Humanoid Dodgeball, a reinforcement learning (RL) framework for agile dynamic obstacle avoidance in humanoids. Our method enables highly dynamic whole-body evasive maneuvers against incoming projectiles through threat-aware reward shaping, while multiple direction-specific adversarial motion priors regularize the learned behaviors toward stable and diverse human-like motions. Extensive simulation and hardware experiments demonstrate that the proposed framework substantially improves dynamic obstacle avoidance performance across diverse obstacle speeds and launch distances, while preserving robust and physically plausible humanoid motions.
Figures & tables
Fig. 1: Our dynamic obstacle avoidance policy deployed on the EngineAI PM01 humanoid. (Top) The robot dodges by stepping forward. (Bottom) The robot dodges by jumping backward.
Fig. 2: Overview of Humanoid Dodgeball . During simulation training, predictive and near-field threat for key bodies are computed to characterize incoming projectile hazards and guide threat-aware reward shaping for whole-body evasion. Direction-specific adversarial motion priors, selected according to the root escape direction, regularize the learned behaviors toward diverse and natural human-like dodge motions. When the threat subsides, recovery rewards, including orientation and capture-point objectives, encourage rapid stabilization. During hardware deployment, the trained actor is executed using obstacle positions estimated from the motion-capture system.
Name
Expression
Weight
Dodge rewards ( T>0 )
Body escape
rescape (Eq. 5 )
40.0
Body distance
rdistance (Eq. 7 )
50.0
Root-body escape direction
rroot−dir (Eq. 8 )
60.0
Recovery rewards in the absence of threats ( T=0 )
Root velocity
exp(−vxyroot22/0.52)
10.0
TABLE I: Reward functions for policy training and their corresponding weights.
Category
Term
Range
Dynamics
Friction coefficient
[0.1,2.0]
Joint Kp,Kd scaling
[0.8,1.2]×
Joint position bias
[−0.05,0.05]rad
Torso COM offset
[−0.05,0.05]m
Sensor noise
Base angular velocity ω
[−0.2,0.2]rad/s
Base roll/pitch (ϕ,θ)
[−0.25,0.25]rad
TABLE II: Domain randomization parameters. All quantities are sampled uniformly within the specified ranges.
Fig. 3: Top: Polar heatmaps of obstacle avoidance success rates across different launch distances and approach directions relative to the robot’s heading. Bottom: Success-rate heatmaps across different obstacle launch speeds and distances. The red contour indicates the 60% success-rate boundary.
Unitree G1
EngineAI PM01
TTH (s)
Method
SR(%)
Trecovery(s)
Pmotor(W)
SR(%)
Trecovery(s)
Pmotor(W)
<0.4
Ours
11.8
1.194±0.356
569.6±124.5
14.2
1.197±0.439
586.3±145.9
REBot [ 14 ]
4.8
1.772±0.468
645.3±93.5
8.1
2.293±0.972
830.9±266.6
PAC-MAN [ 15 ]
5.9
0.825±0.467
316.7±77.2
3.9
1.788±1.382
335.9±139.7
0.4 – 0.8
Ours
64.6
1.438±0.900
587.7±105.7
55.5
1.369±0.666
590.0±129.3
REBot [ 14 ]
10.0
1.735±0.443
618.7±86.1
9.8
2.162±0.575
834.1±228.0
TABLE III: Dodge performance under different time-to-hit ( TTH ) ranges in simulation. Recovery time and average motor power are reported as mean ± standard deviation. Bold indicates the best result for each robot and TTH range.
Fig. 4: Visualization of obstacle collision frequency on the G1 and PM01 humanoids during the simulation evaluation.
Fig. 5: t-SNE visualization of avoidance motions from the evaluation rollouts.
Fig. 6: Ablation studies under different time-to-hit ( TTH ) conditions. (a) Success rate after removing individual reward terms or replacing the direction-specific AMP discriminators with a single discriminator. (b) Fall rate after removing the capture-point recovery reward or using a single AMP discriminator.
Fig. 7: Hardware demonstrations on the EngineAI PM01 humanoid under diverse projectile conditions. (a) A ball aimed toward the robot’s center is avoided by stepping forward. (b) A projectile directed toward the head is avoided through a coordinated forward step and ducking motion. (c) A ball aimed at the right shoulder is avoided by stepping laterally to the left. (d) A projectile targeting the head is avoided through a rightward step combined with a ducking maneuver. These examples demonstrate that the learned policy produces diverse whole-body evasive strategies in response to different obstacle trajectories and incoming directions.
Humanoid locomotion in highly confined environments requires navigating dense environmental obstacles and complex self-collision bounds while maintaining multi-contact dynamic feasibility. Traditional trajectory optimizers frequently struggle in these restricted spaces, as navigating the large collision space with splines on particle abstractions is insufficient and leads to poor local minima. To address this, we propose a three-stage whole-body planning framework that formulates kinematic path planning directly over kinematically reachable rigid-body volumes. By integrating differentiable collision avoidance into a reachability-constrained formulation, our framework synthesizes volume-informed guides that reliably guide a full-order trajectory optimizer over long horizons. We show that these optimized plans serve as high-quality references to train a residual reinforcement learning policy for robust online execution. We validate our approach on the Unitree G1 humanoid across three benchmark testbeds exceeding NIST emergency response standards, achieving restricted confinement ratios (Cr<1.5). Our framework generates feasible trajectories across 12-to-18-second tasks with complex foot and hand contacts where standard baselines fail, while the learned policy successfully tracks these plans under extensive domain randomization in physics simulation.
Carlos Gonzalez, Luis Sentis
Department of Aerospace Engineering and Engineering Mechanics, The University of Texas at Austin, TX 78712, USA.
Robots operating in human-centered environments must safely navigate among multiple dynamic obstacles to avoid collisions with people and surrounding infrastructure. Control barrier functions (CBFs) provide an effective mechanism for safety filtering, and recent CBF-based reinforcement learning (RL) methods embed such safety information into learned policies. However, executing only safety-filtered actions during training can restrict policy exploration, a limitation that becomes particularly consequential in dynamic scenes where safety depends on relative robot-obstacle motion. We propose DODGER, a safety-guided RL framework that directly executes policy-generated actions to drive training rollouts while using CBF-filtered references and constraint violations to shape the policy toward collision-avoidance behavior. We evaluate DODGER through a Dubins-car safety analysis and demonstrate goal-directed navigation among multiple dynamic obstacles in full-order humanoid simulation and real-world humanoid experiments using LiDAR-based perception, without a runtime safety filter.
Sanghyuk Park, Kwanwoo Lee, Taekyung Kim +2
Department of Intelligence and Information, Seoul National University, Republic of Korea · Department of Robotics, University of Michigan, Ann Arbor, MI, USA · Department of Mechanical Engineering, Yonsei University, Republic of Korea +1
Humanoid robots operating in human-centric environments must coordinate the entire body to traverse tight, cluttered spaces while maintaining balance. Existing methods often struggle to jointly consider dynamic stability and collision avoidance from high-dimensional perceptual inputs. Moreover, for humanoids, safe behavior is inherently multimodal: the same obstacle may admit multiple valid traversals, which is difficult to represent with a single deterministic policy. To address these challenges, we propose SWAP (Steerable Whole-body Avoidance Policy), a hierarchical framework that separates collision-aware decision making from stabilization. A high-level, perception-conditioned diffusion policy generates multimodal, collision-aware motion proposals, while a low-level whole-body RL controller tracks these proposals and maintains dynamic balance. To efficiently train the high-level policy, we obtain scalable supervision via hindsight labeling: we roll out the low-level policy under randomized commands, place obstacles only in the space that each executed motion did not sweep, so the motion is collision-free by construction, and re-render the perception the robot would have seen. At deployment, Monte Carlo sampling with simple cost signals steers the diffusion model toward safe, task-aligned behaviors without heavy reward tuning. Experiments in simulation and on a real Unitree G1 humanoid show reliable avoidance with static and dynamic obstacles. Test-time steering further enables customization, improving targeted rewards (e.g., speed, orientation, or posture) by 34% on average while maintaining a high collision avoidance rate. Together, these results support a practical solution to humanoid whole-body collision avoidance with test-time steerability. Project website: https://iscoyizj.github.io/SWAP-steerable-humanoid-oa-website/