Learning Agile Perceptive Traversal of Sparse 3D Structures for Humanoids

Abstract: Traversing sparse 3D structures requires humanoid robots to perceive thin, overhanging geometry while executing agile, accurate whole-body motions. We study this problem through monkey-bar traversal, where the robot must jump to the structure, traverse it through sparse bar interactions, and land safely. For this task, we present a reinforcement-learning-based perceptive control system that operates directly on observations from a head-mounted solid-state lidar. To extract task-relevant geometry from the sparse returns, the policy consumes the raw lidar scan through an attention-based encoder with recurrent memory. This policy is obtained by a phase-scheduled teacher- student pipeline that combines privileged experts for jumping up, brachiating, and jumping down. For transfer to hardware, we model lidar noise, battery-voltage sag, and actuator thermal limits, and equip the humanoid with passive hook end-effectors for robust bar interaction. On hardware, the resulting policy completes the full jump-up->brachiation->jump-down sequence in 14 of 15 trials across three bar configurations and reaches brachiation speeds up to 0.5 m/s. Beyond brachiation, the same perception backbone supports a separately trained policy that ducks beneath thin overhead obstacles with 2 cm cross-sections.

• The paper develops a map-free controller using raw lidar data and learning methods to enable a humanoid robot to navigate sparse 3D structures like monkey bars, achieving a 93% success rate in hardware trials.
• The robot uses a phase-scheduled multi-teacher distillation approach, which consolidates privileged jumping, brachiation, and landing experts into a single policy, demonstrating effective task-specific transfer.
• The hardware platform’s actuation and sensing design incorporates battery-voltage sag and thermal limits, featuring a passive hook mechanism that supports dynamically feasible actions and robust policy responses to uncertain conditions.

Problem setting and contribution

“Learning Agile Perceptive Traversal of Sparse 3D Structures for Humanoids” (2608.29769) addresses a demanding perceptive-control problem: a humanoid must identify centimeter-scale, overhanging structures from sparse onboard observations while generating dynamically feasible whole-body motions. The authors instantiate this problem as monkey-bar traversal, requiring a PM-01 humanoid to jump from the ground to the first bar, brachiate across successive bars, release the final bar, and land safely. Unlike conventional locomotion settings based on elevation maps, voxel grids, or dense depth imagery, the task depends on recognizing thin three-dimensional geometry whose lidar returns are intermittent and highly sensitive to motion and range-discontinuity artifacts.

The paper makes four principal contributions. First, it develops a map-free controller that consumes raw measurements from a head-mounted solid-state lidar through an attention-based encoder and recurrent memory. Second, it proposes a phase-scheduled multi-teacher distillation pipeline that consolidates privileged jumping-up, brachiation, and jumping-down experts into one perceptive policy. Third, it incorporates battery-voltage sag, actuator thermal limits, and sensor-specific lidar corruption into training and sim-to-real transfer. Fourth, it demonstrates both complete monkey-bar traversal and a separate ducking behavior under thin overhead obstacles using the same perception backbone.

The central claim is technically specific: direct raw-lidar control can support contact-accurate, high-dynamic humanoid behaviors on sparse structures without constructing an explicit geometric map. This claim is supported by hardware trials, although the demonstrated policy remains task-specific and is evaluated on a limited set of structures.

Figure 1: The hardware system performs jumping and ducking while the learned attention mechanism concentrates on task-relevant lidar returns.

Hardware and sensing design

The hardware platform is a modified PM-01 humanoid equipped with passive hook end-effectors and a RoboSense E1R solid-state lidar. The hook is fabricated as a stainless-steel plate with an opening accommodating a 60 mm circle, substantially larger than the 1–3 cm radii of the training bars. This geometry deliberately trades grasp specificity for tolerance to placement error. The hook supports loaded contact during swinging, passive capture of bars, and bidirectional traversal. Release is achieved by rotating the hook out of the bar plane through wrist yaw, avoiding the need to lift the robot’s body with the arm actuators.

This mechanical design is integral to the learning system rather than an incidental hardware detail. A passive hook reduces the precision demanded of the policy at capture and reduces the torque required during release. The symmetric construction also permits forward and backward traversal and simplifies simulation through primitive collision geometries. The paper therefore treats end-effector morphology as part of the sim-to-real strategy: robust contact mechanics compensate for unavoidable uncertainty in perception and policy execution.

The lidar provides a nominal $192 \times 144$ scan over a $120^\circ \times 90^\circ$ field of view and includes an integrated IMU. The deployed policy receives a decimated grid of lidar returns at 10 Hz. For brachiation, the scan is cropped to the upper workspace so that the encoder emphasizes the ladder and upcoming bars; ducking retains the full field of view to include both ground and overhead obstacles. A solid-state sensor is selected because electronic scanning is expected to reduce motion-induced distortions relative to mechanically scanning lidar during explosive motion.

The system’s actuation demands are substantial. During a complete sequence, estimated mechanical joint power peaks at 2.11 kW during jump-up, while battery voltage reaches a minimum of 34.7 V. The 18 low-torque actuators approach their limits during voltage sag. These measurements motivate the explicit electrical and thermal models used during training rather than merely serving as post hoc hardware diagnostics.

Figure 2: Estimated mechanical joint power over the jump-up, brachiation, and jump-down phases, including the 2.11 kW jump-up peak.

Privileged teachers and phase-scheduled distillation

The authors separate exploration from perception. Three privileged PPO teachers are trained independently: one for jumping up, one for brachiation, and one for jumping down. Each teacher receives ground-truth bar endpoints and privileged robot state, including contact states, base velocity, battery state, and thermal state. The brachiation teacher is conditioned on goal-position commands, whereas the jump-up and jump-down teachers use contact-based objectives. This specialization allows reinforcement learning to discover distinct contact modes without requiring demonstrations or hand-designed motion primitives.

The deployed student does not observe bar endpoints or the privileged physical states. Instead, the student is trained through a three-stage procedure. Initial behavior cloning uses DAgger-style data aggregation and truncated backpropagation through time. A critic-warmup stage then regresses generalized advantage estimates while continuing actor behavior-cloning updates. Finally, regularized PPO gradually replaces the imitation objective with the task reward through a decaying behavior-cloning anchor.

A key design choice is that the active teacher is scheduled by task phase within a single episode. This differs from maintaining separate policies and switching among them at deployment, as in approaches such as ANYmal Parkour [10.1126/scirobotics.adi7566]. Phase-specific advantage normalization and inverse-return-variance weighting address the heterogeneous reward scales associated with jumping, swinging, and landing. The implication is that the student can optimize a unified recurrent policy while preserving subtask-specific learning signals; however, the method still presupposes a known or reliably inferred phase schedule and does not establish general automatic skill composition.

Figure 3: The architecture combines privileged subtask teachers, a phase scheduler, a recurrent perceptive student, and three-stage multi-teacher distillation.

Raw-lidar perception architecture

The student combines proprioception, commands, and lidar observations. Proprioception includes joint positions and velocities, IMU angular velocity, projected gravity, previous actions, and a four-frame history. Commands specify a planar goal and, where applicable, a jump-down trigger. The policy produces 23 joint-position targets at 50 Hz, tracked by joint-level PD controllers.

The perception encoder adapts the AME-2 attention-based map encoder (Zhang et al., 13 Jan 2026) to raw lidar point measurements. Rather than treating the returns as an unordered point set, it preserves their native two-dimensional scan-grid structure. This allows the attention mechanism to exploit spatial organization while avoiding the memory and resolution costs of dense voxelization. The resulting features are combined with proprioception and commands in a GRU, whose hidden state supplies the temporal integration required when bars appear only intermittently in individual scans.

The recurrent hidden state is also trained with an auxiliary bar-centerline prediction loss. The auxiliary decoder estimates the relative position and orientation of the closest bar, providing a direct geometric training signal during distillation. This is important because the behavior-cloning objective alone supervises actions but does not necessarily force the recurrent representation to encode a geometrically meaningful latent state.

The encoder ablation supports the architectural argument. The attention-based point-cloud encoder with auxiliary supervision obtains a combined behavior-cloning loss of $2.35 \times 10^{-2}$, compared with $2.43 \times 10^{-2}$ without the auxiliary loss, $2.62 \times 10^{-2}$ for a CNN, $2.76 \times 10^{-2}$ for an MLP, and $2.90 \times 10^{-2}$ for a blind proprioceptive policy. It also uses only 13.8k parameters, versus 106.7k for the CNN and 1.31M for the MLP. Thus, the smallest evaluated model both performs best and exploits the scan’s grid structure more effectively than substantially larger generic encoders. The blind-policy result further indicates that proprioception alone is insufficient for the combined traversal task.

The auxiliary loss reduces centerline prediction error from $1.62 \times 10^{-2}$ to $0.71 \times 10^{-2}$ in the reported post hoc probing comparison. Because this probe is not uniformly trained across all ablations, the result demonstrates improved geometric accessibility of the representation but should not be interpreted as a fully controlled measure of perception quality.

Sim-to-real modeling near hardware limits

The transfer strategy models three effects that become consequential during dynamic maneuvers.

The battery model treats all joints as sharing a common battery whose terminal voltage follows a first-order recovery process driven by summed absolute joint torque. Voltage-dependent scaling then reduces motor stall torque and no-load speed before effort clipping. This represents the coupling between whole-body torque demand and the available actuation at every joint. The authors report that omitting voltage sag allowed jump-up to draw enough current to brown out the robot, whereas incorporating the model and its associated reward penalty produced no brownout in hardware trials.

The thermal model assigns a leaky load integrator to 18 low-torque actuators, particularly shoulders and elbows that remain near their torque limits during brachiation. The integrator charges with normalized torque and decays over time. A sustained load ratio of at least 0.2 saturates the integrator under the selected time constants, making prolonged moderate loading penalized rather than only instantaneous overload.

The lidar model is sensor-specific. Hardware inspection identifies ray-cone divergence as the dominant artifact, producing edge bleed at depth discontinuities. The high-throughput training model dithers each ray within a $0.625^\circ$ cone, adds 2 cm Gaussian range noise, injects edge dropout and background mixing, randomizes calibration, corrupts 1% of returns, freezes frames with probability 0.1, and gates ranges to 0.3–1.5 m. The higher-fidelity MuJoCo validation model casts 16 rays per pixel and fuses returns using inverse-square weighting. This distinction between a computationally efficient training model and a more explicit validation model is methodologically appropriate, but the paper does not provide a quantitative sensor-distribution metric showing how closely either model matches the full hardware distribution beyond visual artifact reproduction.

Figure 4: Hardware and simulated lidar observations exhibit comparable edge-bleed and ray-divergence artifacts under the proposed noise models.

Traversal and ducking results

The combined student is evaluated in MuJoCo with a lidar model not used for policy training. Across first-bar heights from 1.65 m to 1.90 m at 0.35 m bar spacing, full-sequence success ranges from 70% to 90%. Brachiation and jump-down each reach 100% in every listed condition, while jump-up is the limiting phase, declining from 90% at intermediate heights to 70% near the upper range. The implication is clear: once the robot is attached to the structure, the learned dynamics are comparatively reliable; the initial ballistic contact remains the principal sensitivity.

Hardware performance is stronger on the tested configurations. Across 15 trials on three ladders, the complete sequence succeeds 14 times, corresponding to a 93% full-sequence success rate. Ladder A produces 9/9 successful sequences, Ladder B 2/2, and Ladder C 3/4. The only failure occurs after successful jump-up, when the hook fails to advance to the next bar. This failure mode is consistent with the paper’s decomposition: perception and control are adequate for initial capture and landing, but contact progression during brachiation remains vulnerable to mechanical or geometric variation. The robot reaches brachiation speeds of up to 0.5 m/s, comparable to the human brachiation speed cited by the authors, although the comparison is not a normalized evaluation of energetic efficiency, morphology, or traversal distance.

Figure 5: The perceptive controller completes jump-up, brachiation, and jump-down across multiple ladder configurations.

The second experiment tests whether the perception backbone is specific to ladder interaction. A separately trained ducking policy reuses the same AME-2 encoder and GRU without architectural modifications, while the privileged teacher uses an upward-directed height-map channel unavailable to the student. In MuJoCo, the policy achieves 100% success across cylindrical obstacles with diameters from 1 to 5 cm and clearances from 1.1 to 1.5 m whenever the robot physically fits. On hardware, it clears randomly oriented wooden bars with a $120^\circ \times 90^\circ$0 cm cross-section in all 10 trials, including a representative passage with 1.2 m minimum clearance.

This result supports a narrower but meaningful claim: the scan-grid attention and recurrent memory are not intrinsically tied to grasping bars. They can encode sparse overhead geometry for a distinct whole-body avoidance behavior. It does not, however, establish zero-shot transfer of the policy itself; ducking requires a separately trained policy and therefore demonstrates reuse of the representation and architecture, not behavioral generalization without retraining.

Figure 6: The ducking policy passes beneath a $120^\circ \times 90^\circ$1 cm wooden slat with 1.2 m minimum clearance.

Limitations and open questions

The evaluation has several explicit constraints. The system consists of a small set of separately trained task policies rather than a general controller capable of selecting or composing skills over arbitrary sparse structures. The ladders vary in height, spacing, width, radius, tilt, and support stability, but the demonstrated geometry remains structurally regular. The paper therefore does not establish robustness to substantially more diverse sparse 3D environments, irregular bar arrangements, occluded structures, or obstacles requiring persistent spatial reconstruction.

The current recurrent memory integrates temporally adjacent observations but does not implement an explicit long-horizon spatial map. It remains open whether the GRU can maintain sufficiently precise geometric state when structures disappear from view for extended intervals or when multiple plausible contact sequences must be disambiguated. Similarly, the phase scheduler simplifies multi-task learning by assigning the active expert according to task progression; the paper leaves open how phase transitions would be inferred under failed contacts, unexpected support motion, or alternative traversal choices.

Finally, the hardware sample is small: 15 complete traversal trials and 10 ducking passes. The 93% traversal success rate is consequential, but confidence intervals and larger-scale randomized testing would be needed to characterize failure probabilities under controlled variations in sensor mounting, battery state, temperature, bar geometry, and structural compliance.

Conclusion

The paper presents an integrated perception, learning, morphology, and transfer approach for humanoid interaction with sparse three-dimensional structures. Its principal technical result is that an attention-based encoder operating directly on sparse solid-state lidar returns, augmented with GRU memory and geometric auxiliary supervision, can support dynamic contact behaviors that are poorly represented by conventional elevation maps. Privileged phase-specific teachers and PPO refinement produce a single traversal policy, while explicit battery, thermal, and lidar models address the failure mechanisms exposed by high-power hardware motion. The system completes the full jump-up–brachiation–jump-down sequence in 14 of 15 hardware trials and clears thin overhead obstacles in all 10 reported passes. The remaining question is whether this design can extend from structured ladders and separately trained skills to long-horizon traversal of substantially more varied sparse geometries.