Source-linked AI summary
DAVIS: A Depth-Only End-to-End Active-Vision Framework for Humanoid Soccer Skills
Jiakang Jin, Yixiao Huo, Pengyuan Wang, Yinan Han, Tingxuan Zhang, Zhuobing Zhao, Xuanxin Zhou, Zhangchen Ye, Enxuan Ruan, Yifei Bao, Jiankun Yang, Chenghao Sun, Wenhao Cui, Xiaoyu Tian, Yiming Li
TL;DR
Humanoid soccer skills must be learned despite a shifting head-mounted view and contact uncertainty, raising the question of whether depth and proprioception can suffice without external scene state. DAVIS trains a depth-only, end-to-end policy and instantiates shooting and dribbling within a shared interface. Noetix E1 execution and additional skill demonstrations support transfer across contact regimes, while depth sensing and task-specific hand-tuning remain limitations.
Problem
Head motion changes the head-mounted view and perception or stance errors can alter contact outcomes, motivating the question of whether soccer skills can use depth, proprioceptive history, and task commands without external scene state.
Source: §I · pp. 1–2Method
DAVIS uses a depth-only end-to-end policy and specifies skills through task objects, commands, geometry, objectives, and curricula while retaining the shared deployment interface.
Source: §I · p. 2, D. Skill Instantiation · p. 4Results
Real-robot execution on the Noetix E1 supports transfer of the depth-only interface across shooting and dribbling contact regimes, alongside additional skill demonstrations.
Source: §V · p. 7Takeaways & Limitations
The results support DAVIS's practical feasibility and applicability across soccer tasks through task-specific training.
Source: §V · p. 7Depth-only sensing inherits a single camera's occlusions, dropouts, and imperfectly simulated depth, while each skill still needs hand-tuned rewards, curricula, and contact conditioning.
Source: §VI · p. 8
Abstract
from arXiv · showhide
Humanoid soccer contact skills require more than producing high-impact foot-ball contacts: the robot must close the loop over perception, approach, alignment, impact, and recovery while its own motion induces substantial viewpoint changes, frequent loss of the ball from view, and uncertain contact outcomes. In this work, we ask a compact yet stricter question: can a humanoid learn soccer contact skills using only a head-mounted depth image, proprioceptive history, and an optional low-dimensional task command, and directly output 25-DoF joint PD targets without extra runtime perception or planning modules? To this end, we propose DAVIS, a depth-only end-to-end framework for humanoid soccer skills that learns visibility-aware auxiliary geometry during training, and combines GT-to-prediction annealing, task curricula, and AMP-style motion priors to smoothly bridge privileged supervision and real deployment. Built on this framework, we instantiate representative soccer contact skills, including goal-directed shooting and directional dribbling, through task-specific definitions of objects, commands, rewards, and curricula, and validate them through simulation, Noetix E1 real-robot experiments, and ablations.
I. INTRODUCTION
Humanoid soccer demands perception, approach, alignment, contact, and recovery despite a changing head-mounted view. DAVIS studies whether depth, proprioceptive history, and task commands can support direct whole-body soccer control.
I. INTRODUCTION
- Moving targets and viewpoint changes make soccer require coordinated perception, approach, alignment, contact, and recovery, with head motion affecting available evidence and contact outcomes.The paper asks whether this can be learned without external scene state, using head-mounted depth, proprioception, and low-dimensional commands.Source: §I · pp. 1–2
- DAVIS maps head-mounted depth, proprioceptive history, and low-dimensional commands directly to 25-DoF joint targets, with active-vision head control.Privileged simulation signals guide training but are not part of the deployed observation-to-control path.Source: §I · p. 2, Figure 2 · p. 3
A. Robot Soccer Skill Learning
- Many existing soccer systems still depend on RGB detection or multisensor pipelines –; DAVIS studies depth-driven chasing, dribbling, and kicking.Source: A. Robot Soccer Skill Learning · p. 2
B. Whole-Body Loco-Manipulation
- DAVIS jointly optimizes active visual tracking, approach, and foot–ball interaction for dynamic ball-centric loco-manipulation.The ball's motion adds demands for suitable approach direction, visibility, and precise foot–ball contact.Source: B. Whole-Body Loco-Manipulation · p. 2
A. Problem Formulation
- The POMDP formulation reserves full simulator state and privileged labels for training, while deployment observes depth, proprioceptive history, and an optional task variable to produce a 25-D action.The actuated head is part of the sensing loop because its action changes the next depth observation.Source: A. Problem Formulation · pp. 2–3
B. Depth-Only End-to-End Structured Soccer RL Framework
DAVIS keeps deployment to a depth-and-proprioception-to-control path, placing task structure in training supervision and regularization. Its auxiliary module predicts object geometry and visibility, with annealing to transition from privileged guidance to depth-based predictions.
B. Depth-Only End-to-End Structured Soccer RL Framework
- The deployment actor uses depth, proprioceptive history, and an optional command to produce 25-DoF joint-target residuals; training adds privileged simulation supervision and motion regularization.Auxiliary geometry and visibility predictions are confidence-gated before entering the actor, and task definitions preserve this deployment interface.Source: B. Depth-Only End-to-End Structured… · p. 3
C. Visibility-Aware Auxiliary Geometry
- The geometry module predicts each task object's 3-D center and visibility from the depth latent, providing object-level supervision without a runtime perception module.The same construction assigns indexed geometry and visibility predictions to multiple objects.Source: C. Visibility-Aware Auxiliary Geometry · pp. 3–4, Figure 3 · p. 4
- Geometry supervision applies only when an object is visible, while visibility receives binary supervision based on camera field-of-view and depth-range constraints.For multi-point objects, visibility is positive when any predefined structural point is valid; the geometry target remains the single object center.Source: C. Visibility-Aware Auxiliary Geometry · p. 4
- GT-to-prediction annealing gradually replaces privileged geometry with predicted geometry, so deployment uses auxiliary features predicted solely from depth.Source: C. Visibility-Aware Auxiliary Geometry · p. 4
D. Skill Instantiation
DAVIS reuses a fixed depth-to-control interface, specifying each soccer skill through task-specific objects, commands, objectives, and curricula. Shooting and directional dribbling illustrate the framework across complementary contact regimes.
D. Skill Instantiation
- A skill definition is T = (O, ct, yt, rt, K), comprising task objects, an optional command, supervised task geometry, a contact objective, and a curriculum.Source: D. Skill Instantiation · p. 4
- New skills retain the depth-to-control path and are introduced by changing task objects, commands, objectives, and curricula.Source: D. Skill Instantiation · p. 4
- Goal-directed shooting and directional dribbling are representative instances used to assess the framework across complementary contact regimes.Source: D. Skill Instantiation · p. 4
1) Goal-Directed Shooting:
Shooting uses ball–goal geometry to target a strike toward the goal, while directional dribbling uses a direction command and sustained ball regulation. Their rewards, visual objectives, and motion control reflect these distinct contact tasks.
1) Goal-Directed Shooting:
- Shooting targets a strike that sends the ball toward the goal, using ball–goal geometry and a behind-ball target to unify chasing, alignment, and contact.Its reward covers run-up, contact, terminal success, visual maintenance, and safety; the curriculum broadens ball placement and approach difficulty.Source: 1) Goal-Directed Shooting · p. 4
- Shooting maintains active vision through head control: the head keeps the ball in view while the body gradually follows the viewing direction.Source: 1) Goal-Directed Shooting · p. 4
2) Directional Dribbling:
- Dribbling follows a commanded direction from a 12-way ring and regulates the ball over time, rather than seeking a single high-speed strike.Its reward covers commanded progress, re-approach, heading, visual maintenance, and out-of-band penalties.Source: 2) Directional Dribbling · p. 4
- Continuous dribbling uses contact-intent-conditioned AMP and phase masks to emphasize different motion states during approach, touch, follow-through, and recovery.Source: 2) Directional Dribbling · p. 4
E. Training and Deployment Pipeline
The section describes a depth-to-control deployment path and evaluates shooting and dribbling in simulation. Results include repeated-S slalom performance, shooting behavior, and a comparison of learned head control with a gaze heuristic.
E. Training and Deployment Pipeline
- The deployment actor maps a single head-mounted depth frame, five-step proprioceptive history, and task variables to 25-DoF joint PD targets.Depth alignment includes distortion correction, cropping to 168 × 80, occlusion, noise, masks, and normalization.Source: E. Training and Deployment Pipeline · p. 4, Figure 4 · p. 5
A. Simulation Experiments
- The shooting and dribbling tasks use separate checkpoints; only the dribbling actor receives a joystick direction command.That command is quantized to a 12-way ring.Source: A. Simulation Experiments · p. 5
- Shooting results demonstrate positional generalization, and the footstep trajectories show a smooth gait with a consistent speed pattern.Each performance region uses over 50 trials.Source: A. Simulation Experiments · p. 5, Figure 6 · p. 6
- 0.65 average success rate across 900 repeated-S slalom trials, with more than half of task configurations above 0.65.The task uses a fixed instruction sequence and requires passing all waypoints in order.Source: Table I · p. 6, A. Simulation Experiments · p. 6
- On highly dynamic dribbling, learned head control remains effective while a heuristic that gazes at the predicted ball fails.For static penalty kicks, the heuristic is comparable to learned active vision.Source: A. Simulation Experiments · p. 6
- At 1 m tolerance, DAVIS reaches target-reaching and obstacle-avoidance success rates of 0.88 and 0.96, versus Dribble Master’s 0.87 and 0.93.DAVIS uses a depth image, while Dribble Master uses detector-derived ball coordinates.Source: A. Simulation Experiments · p. 6
B. Ablation Study
DAVIS is evaluated through simulation ablations and Noetix E1 deployments, with results spanning shooting and dribbling. The authors report that the full method balances ablation metrics across tasks, while identifying perceptual and task-design limits.
B. Ablation Study
- The full method provides the best overall balance across ablation tasks and metrics; active vision has a clear advantage over heuristic gaze in dynamic dribbling.The heuristic controller performs comparably in static-object tasks such as penalty kicks.Source: Table II · p. 7
- Without the progressive curriculum, training takes more iterations to converge and performance is significantly worse than baseline.The ablation activates all reward terms at full weight from the beginning.Source: B. Ablation Study · p. 7
- Masking auxiliary predictions while retaining the architecture causes object localization and contact performance to collapse, with success dropping to zero.The study also tests removing annealing, actor geometry inputs, auxiliary supervision, or visibility gating.Source: B. Ablation Study · p. 7
- Fixing the head joints causes frequent ball loss from view, gate-localization failure, near-zero performance, and a high likelihood of falling.The active-vision components were not ablated independently because they act through the same two head-joint dimensions.Source: B. Ablation Study · p. 7
C. Real-World Deployment
- 26/38, 28/49, and 33/60 shooting trials succeeded in the Easy, Medium, and Hard real-world settings, respectively.Success required scoring with a single kick.Source: C. Real-World Deployment · p. 7
- 22/28, 19/28, and 17/28 real-world dribbling trials succeeded for target reaching, obstacle traversal, and slalom, respectively.Success required remaining upright and finishing with the ball within 1 m of the target.Source: C. Real-World Deployment · p. 7
V. CONCLUSION
- DAVIS keeps a shared depth-to-control interface while task-specific objects, commands, rewards, curricula, and motion-prior conditioning vary across shooting and dribbling.The conclusion presents these as representative one-shot striking and sustained ball-control skills.Source: §V · p. 7
VI. LIMITATIONS AND FUTURE WORK
- Depth-only sensing ties the policy to one head-mounted camera and its occlusions, dropouts, and imperfectly simulated depth; each skill also requires hand-tuned rewards, curricula, and contact conditioning.The authors identify these as limits on real-world accuracy and automatic skill construction.Source: §VI · p. 8
APPENDIX
The appendix specifies reproducibility settings, metrics, and protocols for the shooting and dribbling simulation evaluations. It describes how trials are initialized and how outcomes are measured.
APPENDIX
- Policies are trained in Isaac Lab on eight NVIDIA RTX 4090D GPUs for 20,000 PPO iterations before ONNX export, with the 25-dimensional action vector preserved through deployment.Hardware rates, sensor inputs, and excluded runtime modules are specified in Sec. IV-C.Source: APPENDIX · p. 9
- Table III defines the metrics used in the simulation, ablation, and real-world result tables; protocol-specific metric differences are reported with the corresponding result table.The appendix expands the task-level success definitions with evaluation details.Source: APPENDIX · p. 9, Table III · p. 9
A. Directional Shooting Simulation Protocol
- Shooting simulation evaluates penalty kicks and free kicks under distinct ball and robot initializations, with stratified sampling across position regions to reduce sampling bias.The protocol reports success rate, angular error, and successful contact ratio; Fig. 8 depicts difficulty levels.Source: A. Directional Shooting Simulation… · p. 9, Figure 8 · p. 9
B. Directional Dribbling Simulation Protocol
- The repeated-S dribbling benchmark evaluates nine configurations spanning N = 3, 5, or 7 turns and α = 60°, 90°, or 120°, for 900 trials total.Success rate, velocity error, and time in the control band are measured after the robot first takes over the ball; Fig. 9 visualizes successful trajectories and success bands.Source: B. Directional Dribbling Simulation… · pp. 9–10
C. Dribble Master Protocol Reproduction
The Dribble Master reproduction compares DAVIS simulation results against the published protocol and reports a difference in turning bias. The appendix also specifies simulation-ablation and real-world evaluation procedures.
C. Dribble Master Protocol Reproduction
- The reproduced protocol uses DAVIS’s checkpoint and simulator, aligning the published and reproduced setups in Table V and reporting reaching, obstacle, and turning results in Tables VI and VII.Tables VI and VII identify their sample sizes and evaluation sources: DAVIS simulation uses 50 rounds per row or condition, while the comparison values come from robot trials or MuJoCo rollouts.Source: C. Dribble Master Protocol Reproduction · p. 10, Table V · p. 10, Table VI · p. 11
D. Ablation Experiment Protocol
- Simulation ablations use identical test settings and learning parameters, and no ablation studies are conducted on real-world robots.The protocol removes staged head/body reward scaling, disables auxiliary features and losses, or replaces head-oriented rewards with body alignment rewards; Fig. 10 depicts the active-vision ablation.Source: D. Ablation Experiment Protocol · pp. 10–11, Figure 10 · p. 11
C. Dribble Master Protocol Reproduction
- 7.1% average turning-bias statistic across eight conditions compares with Dribble Master’s 2.2–3.2%; the authors attribute the gap to command-interface differences, not perception.DAVIS commands direction rather than a ball-velocity vector, leaving lateral drift inside the controllable band uncorrected; off-grid 45° conditions track no worse than neighboring 30°/60° controls.Source: C. Dribble Master Protocol Reproduction · p. 10, Table VII · p. 11
E. Real World Experiment Protocol
- Real-world shooting was evaluated from 12 predefined positions within the test area; dribbling tested reaching a target, obstacle traversal, and continuous navigation through sequential target locations.A successful shooting trial required one kick to send the ball completely across the goal line. In the real-world dribbling experiments, both the robot and ball had to finish within 1 m of the target, and trials used the same joystick operator.Source: E. Real World Experiment Protocol · p. 12
F. Observation Space and Policy Architecture
The policy combines depth and proprioceptive history for control, while training-only privileged information supports value learning and auxiliary supervision. Geometry and visibility predictions are gradually introduced into the actor’s inputs during training.
F. Observation Space and Policy Architecture
- The actor receives one aligned depth frame, five steps of proprioceptive history, and an optional task command; a LightDepthEncoder maps depth to a 32-D latent.The shooting and dribbling policies use 81-D and 84-D proprioceptive vectors, respectively, and actor and critic MLPs use [512, 256, 128] hidden layers with ELU activations.Source: F. Observation Space and Policy… · p. 12, Table IX · p. 13
- Privileged quantities such as ball and goal geometry, visibility, and contact states are used for training-time value learning, rewards, or auxiliary supervision, not as deployment actor inputs.The HIM estimator produces a 3-D hidden-state estimate from proprioceptive history; the AMP discriminator masks head-related dimensions because head yaw and pitch serve active visual tracking.Source: F. Observation Space and Policy… · pp. 12–13
G. Auxiliary Geometry and Visibility Losses
- Auxiliary heads predict local 3-D ball and goal positions and object visibility from the shared depth latent; geometry is supervised only on visible frames.Visibility uses a binary label based on whether predefined object points fall within the valid camera field of view and depth range.Source: F. Observation Space and Policy… · p. 12, G. Auxiliary Geometry and Visibility… · p. 13
- The actor’s geometry and visibility inputs transition from ground truth to predictions under a monotone, concave schedule, with the shift frozen when auxiliary-loss EMA exceeds a threshold before a grace update.At deployment the mixture is disabled and auxiliary actor features come from depth-branch predictions; the actor-side visibility gate is detached from policy gradients.Source: G. Auxiliary Geometry and Visibility… · pp. 13–14
H. Goal-Directed Shooting
Goal-directed shooting is trained as a contact task with approach, alignment, contact, success, active-vision, and safety reward terms. Near-ball alignment and contact terms receive greater weight as the robot approaches the ball.
H. Goal-Directed Shooting
- Training uses 1024 parallel environments, 20 s episodes, and a 0.005 s physics step.The control decimation matches the shared training and deployment configuration.Source: H. Goal-Directed Shooting · p. 14
- A smooth distance gate gives near-ball alignment and contact terms larger weights as the robot approaches the ball.Source: H. Goal-Directed Shooting · p. 14
- Shooting uses approach, active-vision, ball–goal alignment, contact, terminal-success, and safety-regularization reward terms.The reward summary is identified as Table XI.Source: H. Goal-Directed Shooting · p. 14
I. Directional Dribbling
The section details directional-dribbling deployment and sim-to-real settings, then describes a close-range powerful-shot variant using the same depth-only interface and auxiliary geometry prediction.
I. Directional Dribbling
- Real and simulated controllers use one aligned 168 × 80 depth frame, range-filtered to 0.3–5.0 m and normalized as 2D/5.0 −1.Proprioceptive history, rather than stacked depth frames, supplies temporal information.Source: §I · p. 14
- Directional-dribbling training randomizes visual observations and robot, ball, ground, and contact parameters to support sim-to-real transfer.The dribbling episode lasts 60 s; training tightens the controllable-band half-width from 0.15 m to 0.10 m and the out-of-band dwell from 6.0 s to 3.5 s.Source: §I · p. 14
J. Powerful Shot.
- The close-range powerful-shot variant keeps the depth-only actor interface and auxiliary ball-geometry prediction, changing its reset distribution, rewards, and motion-prior schedule.Figure 14 presents a real-world example of the powerful shot.Source: J. Powerful Shot · p. 14, Figure 14 · p. 14
- The powerful-shot reward curriculum first emphasizes distance-weighted near-ball leg effort, then adds horizontal ball speed and displacement from the reset position.Reward scales are ramped in sequentially; an AMP kick-motion prior activates when enough parallel environments reach the near-ball zone.Source: J. Powerful Shot · p. 14
K. Dribbling while Avoiding Obstacles.
Obstacle avoidance extends directional dribbling by adding obstacle geometry prediction and steering through the existing reward objective. Ball-loss recovery uses the active-vision policy and staged head- and body-following rewards to search and reacquire.
K. Dribbling while Avoiding Obstacles.
- Obstacle avoidance adds an indexed geometry-and-visibility prediction head for the obstacle, with no privileged obstacle input or change to the actor interface or control loop.The obstacle joins the task object set and uses the shared depth latent.Source: K. Dribbling while Avoiding Obstacles · p. 15
- An artificial potential field deflects the commanded direction through tangential repulsion from the nearest in-path obstacle, weighted by proximity and activated only for obstacles ahead.Direction-aware dribble rewards track the deflected command using the same velocity-direction tracking objective and reward structure.Source: K. Dribbling while Avoiding Obstacles · p. 15
L. Recovery upon Losing the Ball.
- When the ball is lost, penalties discourage prolonged invisibility and moving too fast without reliable visual evidence.The search penalty starts after 0.5 s of continuous invisibility; the invisible-speed penalty limits body-frame forward speed around 0.5 m/s.Source: L. Recovery upon Losing the Ball · pp. 15–16
- Staged head-yaw/pitch and body-following rewards produce closed-loop search and reacquisition while retaining the depth-only actor interface.The head reward starts at 12k steps and ramps over 36k; the body-following reward starts at 48k and ramps over 48k.Source: L. Recovery upon Losing the Ball · p. 16