PaperlayerView exact revision on arXiv

Source-linked AI summary

RAIWAY: A Large-Wheeled Bipedal Robot with Contact-Informed Blind Locomotion

Kyungmin Nam, Jiseong Lee, Donghyuk Choi, Jeongsoo Park, Woojin Kim, Jemin Hwangbo

arXiv:2610.33172v1cs.RO

TL;DR

Wheel locomotion over irregular terrain is governed by wheel–terrain contacts that are difficult to observe directly. RAIWAY distills simulator-derived multi-contact events into a latent representation inferred from proprioception at deployment. The system traverses diverse outdoor terrain and tested stairs up to 15 cm, and its controller outperforms baselines on tested indoor and outdoor tasks.

  • Problem

    Rolling locomotion is contact-dominated, but the decisive wheel–terrain interaction is difficult to instrument and depends on factors beyond rigid-body geometry.

    Source: §I · p. 1

  • Method

    The framework distills simulator-derived multi-contact events into a latent representation that the deployed controller infers from proprioceptive history without contact sensing.

    Source: §I · pp. 1–2

  • Results

    The contact-informed controller achieves markedly higher success rates than baselines on a 16 cm indoor step ascent and outdoor rocky traversal.

    Source: §I · p. 2, B. Ablation Study · p. 6

  • Takeaways & Limitations

    On the tested large-wheeled platform, contact-informed control improves performance across tested conditions while enabling rolling traversal of diverse irregular terrain.

    Source: §V · p. 8

Abstract

from arXiv · show

We present RAIWAY, an in-house wheeled bipedal robot with 20-inch wheels, together with a contact-informed control approach for rolling locomotion across irregular and discontinuous terrain. Such locomotion is fundamentally contact-dominated, governed by local wheel--terrain interaction that is impractical to instrument directly and depends on more than terrain geometry alone. The proposed framework exploits the contact information available in simulation, distilling multi-contact events into a latent representation. At deployment, the controller infers this representation online from proprioception, requiring no contact sensing. In real-world experiments, the proposed system traverses diverse environments including steep slopes, gravel beds, studded surfaces, and rocky terrain, and demonstrates continuous ascent on the tested outdoor staircases with riser heights up to 15~cm. On our platform, the proposed controller further outperforms baselines without the contact-informed mechanism across challenging indoor and outdoor settings. The video can be found at https://youtu.be/Ch0divLg64w.

I. INTRODUCTION

RAIWAY targets continuous rolling over irregular terrain, where wheel–terrain contact is difficult to observe and depends on more than terrain geometry. It combines a 20-inch-wheeled biped with a controller that uses simulator-derived contact information and infers a latent representation from proprioception at deployment.

I. INTRODUCTION

  • Large wheels target discontinuities that can compromise continuous rolling: their larger radius engages obstacle edges at a shallower angle, while tire compliance absorbs impacts.The authors build a 20-inch-wheeled biped to target robust traversal of challenging terrain.Source: §I · pp. 1–2
  • Rolling is contact-dominated: the decisive wheel–terrain state is difficult to observe and depends on impacts, slip, tire compliance, and internal materials.The wheel’s rotating surface is difficult to instrument densely, while external sensing captures surrounding terrain rather than the contact interface.Source: §I · p. 1
  • The controller distills simulator-derived multi-contact events into a latent representation that it infers online from proprioception, without contact sensing.The approach uses a task- and platform-specific representation of wheel–terrain contacts within Regularized Online Adaptation.Source: §I · pp. 1–2
  • RAIWAY traverses steep slopes, gravel, studded surfaces, and rocky terrain, with continuous ascent on tested outdoor stairs up to 15 cm.The figure shows stair risers of 9, 14, and 15 cm.Source: §I · p. 1, Figure 1 · p. 2
  • The contact-informed controller achieves markedly higher success rates than baselines on a 16 cm indoor step ascent and outdoor rocky traversal.The paper concludes that the controller improves performance on its large-wheeled platform across the tested conditions.Source: §I · p. 2
  • The work develops RAIWAY, proposes a contact-informed pipeline, and demonstrates real-world rolling across irregular terrain without external perception or contact sensing.The demonstrated environments include slopes, gravel, studded surfaces, rocky terrain, and stairs.Source: §I · p. 2

II. RELATED WORK

Existing wheeled bipedal platforms commonly address discontinuities with leg-mediated maneuvers rather than general rolling over terrain. This work instead treats wheel–terrain contact as privileged training information and represents its richer multi-contact structure.

A. Existing Wheeled Bipedal Robots and Blind Locomotion

  • Large compliant wheels can negotiate moderate discontinuities mechanically, supporting rolling as the primary traversal mode while legs regulate posture and load transfer.A larger radius reduces the effective approach angle; tire compliance absorbs impacts, maintains traction, and smooths edge engagement.Source: A. Existing Wheeled Bipedal Robots and… · p. 2
  • Prior platforms mainly interrupt or supplement rolling with dedicated leg-mediated behavior, which suits specific obstacle classes but does not extend to general terrain.Examples include planned contact-switching trajectories, stair-climbing mode flags, and contact-triggered leg lifting.Source: A. Existing Wheeled Bipedal Robots and… · p. 2

B. Explicit and Implicit Use of Privileged Information

  • Privileged information can support locomotion through explicit state estimation or implicit adaptation using latent variables.Prior approaches include concurrent policy-estimator training, teacher–student latent reconstruction, and Regularized Online Adaptation.Source: B. Explicit and Implicit Use of… · pp. 2–3
  • Unlike prior locomotion work that uses contact among several privileged inputs, this approach treats contact as the signal of interest and captures richer wheel multi-contact information.Earlier approaches included foot contacts, forces, or normals as privileged inputs for general locomotion.Source: B. Explicit and Implicit Use of… · p. 3

III. METHOD

RAIWAY is a 6-DoF wheeled biped with large wheels and foam tire inserts. Its controller combines proprioceptive encoding, physical-quantity estimation, and contact-informed latent adaptation.

A. RAIWAY: The Large-Wheeled Bipedal Robot

  • RAIWAY is a 6-DoF wheeled biped whose legs each have two pitch joints and one wheel joint, with differential steering and 20-inch off-road tires.The robot weighs approximately 28 kg with its battery and stands 1.14 m tall in its nominal configuration.Source: A. RAIWAY: The Large-Wheeled Bipedal… · p. 3
  • Closed-cell polymer foam tire inserts provide repeatable impact compliance, avoiding excessive rebound at high pressure and rim strikes at low pressure.Their more consistent tire response also benefits sim-to-real transfer by narrowing the gap between simulated and deployment contact dynamics.Source: A. RAIWAY: The Large-Wheeled Bipedal… · p. 3

B. Contact-Informed Policy Learning

  • The controller combines recurrent proprioceptive encoding, estimation of deployable physical quantities, and conditioning on a contact-informed latent representation.It outputs desired positions for four leg joints and angular velocities for two wheel joints, tracked by a low-level PD controller.Source: B. Contact-Informed Policy Learning · p. 3

C. Processing Contact Events in Simulation

The method converts variable simulator wheel contacts into a rotation-invariant, filtered, fixed-size vector for a teacher encoder. A student learns to infer the resulting latent representation from proprioceptive history for deployment.

C. Processing Contact Events in Simulation

  • Raw contact data records each wheel’s contact positions and impulses at simulator substeps, with the number of contacts varying by wheel.The contacts are collected in the world frame for the left and right wheels.Source: C. Processing Contact Events in… · p. 3
  • Variable, temporally discontinuous simulator contacts are converted into a rotation-invariant, filtered, fixed-size vector before teacher-module input.This preprocessing addresses the raw contact set’s variable size and temporal discontinuity.Source: C. Processing Contact Events in… · p. 3
  • A recurrent teacher encoder maps the processed contact-vector sequence to a latent representation, while a student learns to infer it from onboard proprioceptive history for deployment.The caption specifies an LSTM teacher encoder producing a 32-dimensional latent and a student MLP.Source: Figure 3 · p. 4

1) Transforming into Rotation-Invariant Frame:

A wheel-static frame removes wheel-rotation dependence from contact locations, while angular sorting and padding turn filtered contacts into a fixed-size, geometrically ordered representation.

1) Transforming into Rotation-Invariant Frame:

  • The wheel-static frame {Ws} is invariant to wheel rotation, with its axes defined relative to the wheel axle and projected gravity.Contact positions are transformed into this frame and normalized by wheel dimensions.Source: 1) Transforming into… · p. 4

3) Constructing the Contact Information Vector:

  • The asymmetric debounce filter accepts new contacts immediately but commits contact loss only after 25 consecutive substeps (62.5 ms).The delay is based on the 99th-percentile simulated contact-loss durations on smooth terrain and a 5 cm step; longer windows would delay genuine loss further.Source: 3) Constructing the Contact… · p. 4
  • Angular sorting and three-slot padding give recurring contact configurations consistent geometric roles, from bottom-of-wheel rolling to front-side obstacle engagement.Slots are initialized with a default wheel-bottom template distinguishable from active contact.Source: 3) Constructing the Contact… · pp. 4–5, Figure 4 · p. 4

4) Contact-Informed Latent Adaptation:

A teacher latent distilled from high-rate simulated contact information is adapted online from recent proprioceptive observations and commands, then used to fine-tune the deployed policy.

4) Contact-Informed Latent Adaptation:

  • At 400 Hz, the LSTM teacher processes four contact vectors per 100 Hz policy step and outputs a latent comprising four hidden states.The adaptation MLP predicts that latent from the most recent 20 observations and commands.Source: 4) Contact-Informed Latent Adaptation · p. 5
  • Training jointly regularizes the teacher latent to remain predictable from the student prediction, annealing λ from 0 to 1.The actor uses the teacher latent for most training; in the final phase the encoder is frozen and the actor is switched to the student-recovered latent for fine-tuning.Source: 4) Contact-Informed Latent Adaptation · p. 5

D. Training Details

The policy is trained in parallel simulation with domain randomization intended to support transfer to hardware and reward terms favoring continuous wheel-driven rolling.

D. Training Details

  • PPO training uses 200 parallel Raisim environments for 60,000 iterations and takes approximately 42 hours on the reported CPU and GPU.The hardware is an AMD Ryzen 5995WX CPU and NVIDIA GeForce RTX 3080 Ti GPU.Source: D. Training Details · p. 5
  • Domain randomization varies collision-body and terrain parameters to better represent tire dynamics, slip, and rebound observed in real experiments.Generated steps and stairs reach heights up to 15 cm.Source: D. Training Details · p. 5
  • The reward omits gait-shaping and foot-clearance terms, instead penalizing wheel contact loss and excessive slip to favor continuous wheel-driven rolling.Source: D. Training Details · p. 5

IV. RESULTS

Results include simulation ablations of contact-vector components and latent analysis, alongside real-world ablation studies, outdoor deployments, and failure-case analysis.

A. Analysis of Contact-Vector and Contact-Informed Latent

  • By training completion, the teacher latent's ordering by staircase riser height becomes smoother and more continuous under full regularization.The analysis spans riser heights from −20 to +15 cm; integrated-gradient attribution is measured across 16,000 states covering smooth terrain, steps, pipes, and stairs.Source: A. Analysis of Contact-Vector and… · p. 5, Figure 5 · p. 6
  • 74.4% route traversal falls to 50.4% without contact positions and 53.6% without impulses; only three-slot variants retaining both outperform AC + estimator.Removing wheel-level features yields 64.1%, while one or two slots also substantially degrade performance.Source: A. Analysis of Contact-Vector and… · p. 5, Table III · p. 6

B. Ablation Study

The ablations compare contact-informed control with baseline controllers on a 16 cm step and rocky traversal. The full controller combines the latent and concurrent estimator, with the latent contributing most to rocky traversal and the estimator complementing step climbing.

B. Ablation Study

  • The step experiment assessed 20 trials on a 16 cm step, while rocky traversal covered a 20 m round trip averaged across five trials.Step outcomes classified topples as failures and completed ascents with excessive yaw deviation or peak pitch velocity as unstable; rocky-route distance was sampled at ten checkpoints.Source: B. Ablation Study · p. 6
  • Plain AC and AC + estimator struggled across both experiments, while Ours and its reduced variants exceeded both baselines.AC + estimator improved step climbing but performed worse on rocky terrain; simulation analysis found larger rocky-terrain errors for eight of ten estimator outputs, including all six wheel-specific quantities.Source: B. Ablation Study · p. 6
  • Rocky traversal rose from 39% with AC + estimator to 83% with Ours w/o Estimator; Ours with both mechanisms achieved 100% step success and the highest rocky traversal distance.The results indicate complementary roles: the latent addresses distributed irregular contacts, while the estimator supports low-variability single discontinuities.Source: B. Ablation Study · p. 6
  • During a 16 cm step ascent, wheel rotation exceeded the geometrically required 68.3° by 52.6° by T5, associated with transient contact loss.The figure tracks left-leg joint velocities and torques, latent distance, representative contact location, and wheel rotation; the latent distance rises at initial impact and trends back toward its standing reference after re-engagement.Source: B. Ablation Study · pp. 6–7, Figure 6 · p. 6

C. Outdoor Deployment across Diverse Terrains

A single policy traversed outdoor environments without retuning, terrain-specific switching, or external perception. It maintained rolling across varied contact conditions and ascended tested staircases with risers up to 15 cm, though success varied by staircase.

C. Outdoor Deployment across Diverse Terrains

  • The single deployed policy maintained continuous rolling at high commanded speeds across four environments with repeated impacts, shifting slope traction, and yielding gravel.These tests probed terrain without severe discontinuities, including rapid stair descent, steep slopes, and gravel.Source: C. Outdoor Deployment across Diverse… · p. 7
  • On studded and rocky terrain, contact locations shifted unpredictably; estimator and latent traces responded to abrupt, short-duration changes.During rock clearance, the left-wheel slip estimate changed from negative during wheel spin to positive during clearance, while latent components separated slightly earlier; lateral body motion on the studded slope appeared in the lateral-foot-velocity estimate.Source: C. Outdoor Deployment across Diverse… · p. 7
  • The robot completed 11 consecutive round trips on the 14 cm staircase, while it succeeded in 3 of 10 trials on the 15 cm staircase.The 14 cm staircase had a mean ascent speed of 0.28 m/s; the 15 cm staircase had a mean of 0.42 m/s in successful trials. On the 9 cm staircase, six trials yielded a mean ascent speed of 0.51 m/s.Source: C. Outdoor Deployment across Diverse… · p. 7

D. Failure Cases

The failure cases expose limits in tire-contact modeling and wheel-rebound transfer from simulation to hardware. The conclusion reports outdoor traversal through 15 cm staircases and identifies exteroceptive perception as future work for anticipatory awareness and faster traversal of more complex discontinuities.

D. Failure Cases

  • The robot reached the top in 3 of 10 trials on the 15 cm granite staircase, despite succeeding in all 20 trials on the indoor 16 cm wooden step.Rounded, uneven granite provided no corner for tread blocks to catch, and intermittent wheel slip led to forward toppling; the rigid-contact simulator does not model these surface variations or tire engagement and compliance.Source: D. Failure Cases · pp. 7–8
  • Strong wheel rebound sometimes triggered action and joint-torque oscillations that persisted into unfavorable whole-body motion and falls.The measured tire restitution was about 0.61, but training sampled only up to 0.3 because larger values made the policy overly conservative, leaving hardware rebound events outside the training distribution.Source: D. Failure Cases · p. 8

V. CONCLUSION

  • RAIWAY traversed tested outdoor staircases with risers up to 15 cm using contact-responsive locomotion without external perception or contact sensing.The conclusion frames the benefits of contact-informed control on a large-wheeled platform where obstacle traversal is driven primarily by wheel rolling.Source: §V · p. 8
  • Future work will add exteroceptive perception to combine anticipatory terrain awareness with contact-responsive behavior for faster traversal of more complex discontinuities.Source: §V · p. 8
Loading 2610.33172v1…