PaperlayerView exact revision on arXiv

Source-linked AI summary

Stay Seated: Learning Omnidirectional Humanoid Locomotion on a Passive Mobile Chair with Casters

Kango Yanagida, Kazuki Miyazawa, Takato Horii

arXiv:2608.28090v1cs.RO

TL;DR

Humanoid seated locomotion requires learning omnidirectional movement while maintaining unfixed pelvis–seat contact and using intermittent foot–floor propulsion. The paper extends standing velocity-tracking reinforcement learning with a passive-chair model and privileged critic observations, and its actor uses only proprioception and velocity commands. Policies tracked commands across nearly all 20-s rollouts, some outperformed Standing in velocity tracking, and the learned behavior transferred zero-shot to a Unitree G1.

  • Problem

    The paper addresses omnidirectional seated locomotion on a passive mobile chair as a first step toward seated loco-manipulation.

  • Method

    The study minimally extends a standing velocity-tracking environment with a passive-chair model, seated-state rewards, critic-only chair observations, and SY, FS, and CC training components.

  • Results

    Policies tracked omnidirectional commands through nearly all 20-s random-command rollouts, with SY+CC and SY+FS+CC able to outperform Standing in velocity tracking and transfer zero-shot to a Unitree G1.

  • Takeaways & Limitations

    FS reduced CoT but increased tracking error, while adding SY or CC avoided some FS-only stationary local optima without retuning FS; SY also promoted bilateral leg symmetry.

  • Takeaways & Limitations

    The study is limited to one chair, one floor condition, and one command range, and energy savings were not quantified on the physical robot.

Abstract

from arXiv · show

Humanoid robots with quasi-direct-drive actuators continuously generate joint torque while standing, whereas seated humans delegate weight support to chairs during desk work. As a first step toward seated loco-manipulation, we study omnidirectional seated locomotion on a passive mobile chair, requiring unfixed pelvis-seat contact and intermittent foot-floor propulsion of the robot-chair system. We extend a standard standing velocity-tracking environment with a passive-chair model, seated-state rewards, critic-only chair observations, and task-tailored contact settings. The policy is learned without motion-imitation rewards; its actor uses only proprioception and velocity commands, without contact sensing or chair states. In random-command evaluation, the policies tracked omnidirectional commands through nearly all 20-s rollouts, and the best seated policies could outperform the Standing policy in velocity tracking. Across four training seeds, a $2^3$ full-factorial comparison of symmetry regularization (SY), foot-slip regularization (FS), and command curriculum (CC) showed that FS reduced CoT but increased tracking error and that some FS-only policies converged to stationary local optima. Combining FS with either SY or CC avoided this failure without retuning FS, while SY improved bilateral leg symmetry during longitudinal motion. Direction-resolved analysis showed CoT ordered backward $<$ lateral $\ll$ forward, with planted-leg extension in backward and lateral motion and knee flexion following heel contact in forward motion. The learned policy achieved zero-shot sim-to-real transfer to a Unitree G1 and generated omnidirectional seated locomotion.

I. INTRODUCTION

This study takes seated humanoid locomotion on a passive mobile chair as a first step toward seated loco-manipulation. It investigates omnidirectional velocity tracking with unfixed pelvis–seat contact and direct foot–floor propulsion, using a minimal extension of standing locomotion learning.

  • Motivation: Seated behavior could complement standing behavior during extended tabletop tasks by delegating body-weight support to a chair with casters.Humans typically perform desk work while seated, while casters permit changes in working position and viewpoint without standing.
  • Task formulation: Seated locomotion combines unfixed pelvis–seat contact with direct foot–floor propulsion of the robot–chair system.This contact configuration differs from prior wheeled and sliding-platform tasks, where contact is often constrained and ground interaction occurs primarily through the device.
  • Approach: The task is learned by minimally extending a standard standing velocity-tracking environment, without motion-reference rewards or a new learning algorithm.The study examines symmetry regularization, foot-slip regularization, and command curriculum through a full-factorial design.
  • Contributions: 2^3 full-factorial comparison showed that foot-slip regularization alone could produce stationary local optima, whereas combining it with symmetry regularization or command curriculum avoided this failure.The comparison also found that symmetry reliably promoted bilateral leg behavior in seated locomotion.
  • Contributions: The learned actor used proprioception and velocity commands without contact sensing or chair states, and transferred to a physical robot.The study targeted omnidirectional seated locomotion rather than motion imitation.

II. RELATED WORK

Prior humanoid locomotion research has used wheeled and sliding platforms, reference motions, and analytical gait generators, but seated locomotion introduces a distinct contact and control problem. This study learns omnidirectional seated locomotion without motion references or gait templates.

  • Humanoid Locomotion with Wheeled and Sliding Platforms: Prior work studied roller and inline skates, skateboards, skis, and two-wheeled scooters for humanoid mobility.These systems account for coupled humanoid–device dynamics and nonholonomic constraints from wheels or sliding surfaces.
  • Humanoid Locomotion with Wheeled and Sliding Platforms: Many existing tasks constrain body–device contact in advance and rely primarily on the device for ground interaction.Skateboarding is an exception involving direct foot–ground propulsion, but pushing and steering are treated as distinct phases in HUSKY.
  • Task distinction: Seated locomotion jointly requires unfixed pelvis–seat contact, direct foot–floor propulsion, and omnidirectional translational and yaw-command tracking.The paper does not rank this task against prior platform locomotion because the contact topology and control objective differ.
  • Reference-Based Humanoid Locomotion Learning: Unlike human-constrained seated walking, this work targets a quasi-direct-drive humanoid and learns omnidirectional seated locomotion without motion references.The actor receives only proprioception and velocity commands, excluding foot–floor and pelvis–seat contact sensing and chair states.
  • Reference-Based Humanoid Locomotion Learning: Reference-based DRL would require synchronized capture and retargeting of human, chair, and contact motions across directions and speeds.Seated locomotion must reconcile foot–floor and body–object contacts dynamically at the same time.
  • Analytical Gait Generation: Analytical gait generators have primarily addressed standing locomotion, leaving appropriate contact cycles for omnidirectional seated locomotion unclear.The study therefore uses task-specific rewards and regularization terms rather than reference motions or gait templates.
  • Framework: The learning framework extends a standard standing velocity-tracking environment with a passive-chair model, seated-state rewards, critic-only chair states, and pelvis–seat contact settings.The framework also includes the robot–chair system, observation and action spaces, reward, and training conditions.

A. Seated-Locomotion Problem Formulation

The task is formulated as a Markov decision process in which a seated humanoid tracks torso-frame velocity commands while maintaining unfixed pelvis–seat contact and propelling the chair through intermittent foot–floor contact.

  • A. Seated-Locomotion Problem Formulation: The seated-locomotion task is modeled as a discrete-time Markov decision process M = (S, A, P, r, γ).The state contains robot and chair simulation states, contact states, and the velocity command.
  • A. Seated-Locomotion Problem Formulation: At each time step, the actor constructs an observation from the state and samples an action from πϕ(at | oactor_t).The policy parameters are optimized to maximize expected discounted return.
  • A. Seated-Locomotion Problem Formulation: The policy is conditioned on torso-frame forward and lateral translational velocities and yaw rate.These commands specify the desired seated locomotion behavior.
  • A. Seated-Locomotion Problem Formulation: The robot must maintain pelvis–seat contact while tracking commands and propelling the robot–chair system through intermittent foot–floor contact.The robot is not rigidly attached to the chair.
  • B. Robot–Chair System and Contact Modeling: The task uses a 29-degree-of-freedom Unitree G1 with a five-caster mobile chair and rotating seat.Each caster is modeled as a sphere connected to a chair leg through a ball joint.
  • B. Robot–Chair System and Contact Modeling: Each episode lasts 20 s, or 1,000 control steps, and terminates after excessive torso or chair tilt or one second of lost pelvis–seat contact.The tilt threshold is more than 70° from the gravity direction.
  • B. Robot–Chair System and Contact Modeling: A refined pelvis collision mesh and elliptic friction cone suppress excessive pelvis–seat penetration and unnatural lateral slip.The simulation uses a 0.005 s time step with control decimation of four and specified MuJoCo solver settings.

C. Observation and Action Spaces

The asymmetric actor–critic separates hardware-available actor inputs from privileged simulator information used by the critic. The actor outputs normalized joint-position commands that are converted into PD targets around the initial seated posture.

  • Observation spaces: The asymmetric actor–critic architecture gives the actor only hardware-available proprioception and velocity commands.The actor observation excludes pelvis–seat contact, foot–floor contact, and chair states.
  • Observation spaces: The critic additionally receives simulator-only robot, chair, and contact information during training.The actor observation has 96 dimensions, while the critic observation has 217 dimensions.
  • Action space: The actor outputs a normalized 29-dimensional joint-position command centered on the initial seated posture.These outputs are converted into proportional–derivative controller targets.
  • Training settings: The training settings are summarized in Table II, while Table III specifies domain-randomization ranges.The framework uses these settings for simulation training and transfer preparation.
  • Action space: qtarget,t = qdefault + s⊙at converts the normalized action into a joint-position target using a default posture and joint-wise scale.The operator ⊙ denotes element-wise multiplication.

D. Reward Design

The reward combines velocity tracking, posture stabilization, seated-state maintenance, foot-motion penalties, and motion regularization. Seated-state and foot terms encode chair contact, posture, clearance, slip, and landing behavior.

  • Reward composition: The reward combines velocity-command tracking, posture stabilization, seated-state maintenance, foot motion, and motion regularization.These components are summed at each time step.
  • Task and posture rewards: The task reward encourages target translational and yaw-velocity tracking, while the posture reward keeps the torso and chair upright and limits joint deviation.
  • Seated-state reward: The seated-state reward uses pelvis position, chair position, and pelvis–seat contact state to maintain the seated configuration.Chair velocity is not directly rewarded; chair motion arises through robot velocity tracking and robot–chair contact.
  • Foot penalties: The foot penalty includes foot clearance, tangential foot velocity during contact, and landing impact.The factorial comparison separately varies the tangential foot-slip penalty.
  • Motion regularization: Regularization penalizes torso angular velocity, whole-body angular momentum, self-collision, joint-limit violation, joint acceleration, and action-rate variation.The reward also penalizes deviation from the initial seated posture under a zero command.

E. Training Setup and Factorial Conditions

Training uses PPO conditions that vary symmetry regularization, foot-slip regularization, and command curriculum in a 2^3 full-factorial design. A standing policy provides a separate comparison condition under the same velocity commands.

  • Training setup: Each condition is trained independently with PPO using domain randomization, periodic velocity disturbances, and actor-only observation noise.Model parameters are sampled at environment startup, disturbances occur every 2–5 s, and noise is added only to actor inputs.
  • Training components: The factorial design varies symmetry regularization, foot-slip penalty, and command curriculum.SY uses mirrored augmentation and a mirror-loss coefficient λmir = 1.0; FS is enabled with wslip = 0.25; CC expands translational command ranges during training.
  • Factorial conditions: The 2^3 design produces eight conditions: Baseline, SY, FS, CC, SY+FS, SY+CC, FS+CC, and SY+FS+CC.All conditions otherwise share the training setup.
  • Comparison condition: The Standing comparison tracks the same velocity commands using a separate mjlab standing-locomotion environment and modified actor observations.Its default arm posture matches the seated policies.

IV. EVALUATION AND RESULTS

Seated policies maintained the seated state across nearly all random-command rollouts and achieved strong omnidirectional tracking. SY+CC was most consistent, while combining FS with SY or CC prevented the stationary failure seen with FS alone.

  • Random-command evaluation: 1,000 20-s rollouts per condition evaluated velocity-tracking RMSE, timeout success, seat-relative displacement, and torso tilt under resampled random commands.Commands were resampled after 10 s while training-time randomization, disturbances, and observation noise remained active.
  • Overall tracking and seated state: At least 99.45% of rollouts across all eight training conditions completed while maintaining small seat-relative displacement and torso tilt.Timeout success indicates seated-state maintenance but does not by itself establish locomotion.
  • Overall tracking and seated state: SY+CC achieved the lowest seed-averaged translational tracking errors and highest timeout success, while SY+FS+CC delivered similar tracking.Both conditions were numerically lower than Standing on all three tracking RMSEs, though the comparison uses different training environments and rewards.
  • Fixed-command evaluation: SY+CC and SY+FS+CC maintained large command areas across error thresholds, whereas FS had a smaller command area and greater variation across seeds in the vx–vy plane.Differences were smaller in the vx–ωz plane, where SY+FS had the largest command area at low error thresholds.
  • Training-component effects: FS alone caused stationary local optima under some diagonal-forward commands, while adding SY or CC avoided this failure.The failure arose because the foot-slip penalty suppressed foot motions needed for propulsion.

B. Direction- and Speed-Resolved Gait Characteristics

Direction and speed strongly shaped seated gait mechanics, energy efficiency, and tracking. Symmetry regularization balanced bilateral longitudinal contacts, while foot-slip regularization traded lower CoT for higher tracking error.

  • Longitudinal motion: During longitudinal motion, SY+CC and SY+FS+CC yielded nearly identical left and right contact fractions and forces, unlike CC.CC produced left–right differences in contact fraction and force.
  • Lateral motion: During lateral motion, the leg opposite the movement direction had the larger single-foot contact fraction and contact force, with the pattern reversing between leftward and rightward motion.This identifies the opposite leg as the primary propulsive leg.
  • Energy efficiency: At 1.0 m/s, CoT followed backward < lateral ≪ forward for both SY+CC and SY+FS+CC.Forward CoT was roughly twice the lateral value.
  • Energy efficiency: SY+FS+CC reduced CoT relative to SY+CC for several fixed commands, particularly forward motion, whereas SY+CC generally gave lower tracking RMSE.This indicates a trade-off between tracking and FS-induced motion regularization.
  • Gait mechanics: Backward and lateral motion used planted-leg knee extension for propulsion, whereas forward motion followed heel-first contact with knee flexion.The backward pattern kept the heel firmly planted; forward motion used the opposite knee action.

C. Sim-to-Real Deployment

The seated-locomotion policy transferred zero-shot to a physical Unitree G1 and generated multiple seated motion directions without contact sensing or chair states. Training components produced distinct energy, tracking, symmetry, and failure-mode effects, while direction-dependent contact strategies shaped locomotion cost.

  • Physical deployment: Zero-shot deployment on a Unitree G1 generated forward, backward, lateral, and turning motions while maintaining a seated posture.The actor used only proprioception and velocity commands, without contact sensing or chair states.
  • Physical deployment: Nearly all 20-s random-command rollouts tracked omnidirectional velocity commands, and SY+CC and SY+FS+CC could outperform the Standing policy in velocity tracking.The physical-robot trials provide qualitative transfer evidence rather than quantified tracking or robustness measurements.
  • Training effects: FS reduced CoT but increased tracking error, exposing a trade-off between energy efficiency and command-tracking performance.Some FS-only seeds converged to stationary local optima that ignored diagonal-forward commands.
  • Training effects: Adding SY or CC to FS avoided the stationary local optimum across evaluated seeds without retuning the FS weight.CC may have helped by teaching required foot trajectories first under easier low-speed commands.
  • Training effects: SY improved left–right leg symmetry during longitudinal motion, where seated locomotion could otherwise use asymmetric leg solutions.Standing locomotion was already nearly symmetric without SY because both legs support and propel the body.
  • Direction-dependent behavior: At high speed, CoT followed backward < lateral ≪ forward, with planted-leg extension in backward and lateral motion but knee flexion after heel contact forward.Fast forward motion may approach the friction limit, helping explain its higher CoT and tracking error.
  • Energy comparison: A preliminary stationary comparison measured 111.06 W standing and 101.25 W seated, but one trial per posture prevents establishing a general energy-saving benefit.The study was also limited to one chair, floor condition, and command range.

VI. CONCLUSION

The study learned omnidirectional seated locomotion on a passive mobile chair without motion-imitation rewards and transferred the policy zero-shot to a physical Unitree G1. It also identified training trade-offs, symmetry effects, and direction-specific locomotion strategies.

  • Conclusion: The learned policy maintained unfixed pelvis–seat contact, propelled the robot–chair system through intermittent foot–floor contact, and tracked velocity commands in nearly all 20-s random-command rollouts.The actor used no motion-imitation rewards, contact sensing, or chair states.
  • Conclusion: SY+CC and SY+FS+CC could outperform the Standing policy in velocity tracking, while FS reduced CoT but increased tracking error.Some FS-only seeds converged to stationary local optima; adding SY or CC avoided this failure without retuning FS.
  • Conclusion: SY promoted bilateral leg symmetry during longitudinal seated locomotion, where asymmetric solutions were also feasible.This distinguishes seated locomotion from standing locomotion, which naturally coordinates both legs for support and propulsion.
  • Conclusion: At high speed, CoT followed backward < lateral ≪ forward; backward and lateral motion used planted-leg extension, whereas forward motion used knee flexion following heel contact.These direction-specific contact strategies accompanied the observed cost ordering.
  • Conclusion: The contact-sensorless policy transferred zero-shot to a physical Unitree G1.Deployment generated omnidirectional seated locomotion without additional fine-tuning.

APPENDIX STATIONARY POWER COMPARISON

The appendix presents a preliminary stationary battery-side power comparison between standing and seated postures. Seated operation used less measured power in one trial, but the extrapolation does not establish a general energy-saving benefit.

  • Extrapolation: +22 min of operation was estimated for seated posture using constant-power extrapolation for a 421.2 Wh battery.The corresponding extrapolated durations were 3.79 h standing and 4.16 h seated.
Loading 2608.28090v1…