💻 computer science

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

This paper presents a reinforcement learning framework that enables a humanoid robot to achieve robust, zero-shot sim-to-real omnidirectional seated locomotion on a passive mobile chair by extending standard velocity-tracking environments with specialized rewards and contact settings, while demonstrating that combining foot-slip regularization with symmetry or curriculum learning effectively balances energy efficiency and tracking performance.

Original authors: Kango Yanagida, Kazuki Miyazawa, Takato Horii

Published 2026-08-31

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Kango Yanagida, Kazuki Miyazawa, Takato Horii

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a robot that can sit down and work at a desk just like a human. For years, engineers have taught robots to walk on two legs, but standing still requires constant energy; the robot's motors must work hard just to keep its balance against gravity, generating heat and draining power. Humans, however, have a simpler solution for long tasks: we sit. By resting our weight on a chair, we free our muscles and can move around the office without ever standing up. This simple observation sparked a new question for robotics researchers: can a humanoid robot learn to move around while seated, using a wheeled chair and its own feet to push itself forward, backward, and sideways?

A team of researchers set out to answer this by teaching a robot to navigate while sitting on a passive mobile chair. Unlike a self-driving wheelchair that moves on its own, this chair has no motor; it is a simple seat on wheels. To move, the robot must use its feet to push against the floor, propelling both itself and the chair in any direction. The challenge is complex because the robot cannot simply lock its body to the seat. It must maintain a delicate, shifting balance, keeping its hips in contact with the chair while its feet find purchase on the ground to generate movement. If the robot loses contact with the seat or slips too much, the task fails.

To solve this, the researchers used a method called deep reinforcement learning, a process where a computer program learns by trial and error in a virtual world. They started with a standard training environment designed for standing robots and modified it to include the seated scenario. They did not show the robot videos of humans moving or give it a pre-written set of instructions on how to walk. Instead, they simply told the robot what speed and direction it should be moving, and let it figure out the rest. The robot's "brain" received only basic information about its own body position and the command it was given, without any special sensors to feel the chair or the floor.

The researchers tested several different ways to guide the learning process. They wanted to see if adding specific rules about how the robot should move its feet or how it should treat its left and right legs would help. One rule penalized the robot if its feet slipped on the floor, hoping this would encourage smoother movement. Another rule encouraged the robot to use its left and right legs equally, while a third rule started the robot with easy, slow commands before gradually increasing the difficulty.

The results revealed a surprising truth about how the robot learned. When the researchers only used the rule against foot slipping, the robot often failed in a strange way. Instead of learning to push itself forward, some versions of the robot simply gave up and stayed perfectly still, ignoring commands to move diagonally. It had found a "local trap," a solution that technically avoided slipping but failed the actual task. However, when the researchers combined the no-slip rule with either the symmetry rule or the difficulty-gradual rule, the robot learned successfully. It discovered how to propel itself in all directions without getting stuck.

The most effective approach combined a gradual increase in speed with a rule that encouraged balanced leg movement. This combination allowed the robot to track commands with high precision, often performing better than a similar robot trained to stand and walk. The researchers analyzed exactly how the robot moved and found that the direction of travel changed the way it used its legs. When moving backward or sideways, the robot would plant one foot firmly and extend its knee to push the chair away. When moving forward, it did the opposite: it would touch the ground with its heel first and then bend its knee to pull itself and the chair toward that foot.

This difference in movement had a direct impact on energy efficiency. The robot was most efficient when moving backward, followed by moving sideways. Moving forward was the most energy-intensive task, requiring roughly twice as much effort as moving sideways at high speeds. This was likely because the forward motion relied on bending the knee after the foot touched down, a movement that generated more friction and required more force to maintain speed.

The study concluded with a successful test in the real world. The researchers took the computer program they had trained in simulation and installed it directly onto a physical robot without any further adjustments. The robot, sitting on a real chair with wheels, immediately began to move forward, backward, and sideways, and even turned, just as it had in the virtual training. It managed to stay seated and follow commands using only its internal sensors, proving that a robot can learn to navigate a complex, seated environment without needing to see the chair or feel the floor. This work suggests that robots could soon join us at our desks, moving around to grab tools or reposition themselves while remaining seated, saving energy and mimicking a very human way of working.

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

1. Problem Formulation

This study addresses the challenge of seated loco-manipulation for humanoid robots, specifically focusing on omnidirectional locomotion while seated on a passive mobile chair equipped with casters. Unlike traditional deep reinforcement learning (DRL) for humanoid locomotion, which assumes a standing posture where the robot must continuously generate joint torque to support its weight, this task involves a robot delegating weight support to a chair.

The core difficulty lies in the unique contact topology:

Unfixed Pelvis–Seat Contact: The robot's pelvis is not rigidly attached to the chair; it must maintain contact without slipping off while the chair moves.
Direct Foot–Floor Propulsion: The robot-propelled system relies on intermittent foot–floor contact to move both the robot and the passive chair.
Omnidirectional Control: The robot must track translational velocity commands ( $v_x, v_y$ ) and yaw rate ( $\omega_z$ ) in all directions without a dedicated motion reference or human demonstration.

The goal is to learn a control policy that enables the robot to move and manipulate objects while remaining seated, serving as a first step toward efficient seated loco-manipulation for extended tabletop tasks.

2. Methodology

The authors extend a standard standing velocity-tracking environment to accommodate seated locomotion using Proximal Policy Optimization (PPO) within the mjlab simulation framework.

A. System and Modeling

Robot: A 29-degree-of-freedom Unitree G1 humanoid.
Chair: A passive mobile chair with five casters and a rotating seat.
Contact Modeling: The pelvis collision mesh is refined to match the visual geometry to prevent unnatural lateral slipping. The contact solver uses an elliptic friction cone instead of a pyramidal one to better handle the pelvis–seat interaction.
Observation Space:
- Actor: Uses only proprioception (joint positions/velocities, torso angular velocity, gravity projection) and velocity commands. Crucially, it excludes contact sensing (foot-floor, pelvis-seat) and chair states to facilitate sim-to-real transfer.
- Critic: Receives the actor's observation plus privileged information (simulator-only states including chair pose, contact forces, and foot heights).

B. Reward Design

The reward function combines several terms:

Task Reward ( $r_{task}$ ): Encourages tracking of target translational and yaw velocities.
Posture Reward ( $r_{posture}$ ): Penalizes excessive torso/chair tilt and joint deviation from the initial seated posture.
Seated-State Reward ( $r_{seat}$ ): Encourages maintaining the pelvis position relative to the chair and ensuring pelvis–seat contact.
Penalties: Includes penalties for foot motion, self-collision, joint limits, and specific regularization terms.

C. Training Components and Factorial Study

The study employs a $2^3$ full-factorial design to evaluate the impact of three specific training components:

Symmetry Regularization (SY): Applies left-right symmetry data augmentation and a mirror-loss term to the actor loss.
Foot-Slip Penalty (FS): Penalizes tangential foot velocity during contact to reduce energy consumption (Cost of Transport, CoT).
Command Curriculum (CC): A schedule that starts training with lower velocity command ranges ( $\pm 0.5$ m/s) before expanding to the full range ( $\pm 1.0$ m/s).

The authors compare eight conditions (all combinations of SY, FS, and CC) against a baseline and a standing locomotion policy.

3. Key Contributions

Formulation of Seated Locomotion: The paper defines seated locomotion on a passive chair as an omnidirectional velocity-tracking problem with unfixed body–tool contact and direct foot–floor propulsion, learned without motion references or human demonstrations.
Zero-Shot Sim-to-Real Transfer: The learned policy, which relies solely on proprioception and velocity commands (no contact sensing or chair state estimation), was successfully transferred to a physical Unitree G1 robot without fine-tuning.
Analysis of Training Components:
- FS Limitations: The study reveals that Foot-Slip (FS) regularization alone can cause policies to converge to stationary local optima, where the robot remains still to avoid slip penalties, failing to track diagonal-forward commands.
- Mitigation: Combining FS with either Symmetry Regularization (SY) or Command Curriculum (CC) prevents this failure without requiring retuning of the FS weight.
- Symmetry Necessity: While standing locomotion naturally yields symmetric leg usage, seated locomotion allows for asymmetric solutions (one leg dominating propulsion). SY is shown to be critical for enforcing bilateral symmetry in seated tasks, which is desirable for hardware longevity and stability.

4. Results and Analysis

A. Tracking Performance

Success Rate: All trained policies maintained the seated state in over 99% of 20-second random-command rollouts.
Tracking Accuracy: The best-performing conditions (SY+CC and SY+FS+CC) achieved translational tracking errors comparable to or slightly better than the standing policy.
FS Impact: Policies using FS alone showed high variance and often failed to track diagonal commands due to the stationary local optimum. When combined with SY or CC, tracking performance improved significantly.

B. Gait Characteristics and Energy Efficiency

Direction-Dependent CoT: The Cost of Transport (CoT) followed the order: Backward < Lateral $\ll$ Forward.
- Backward/Lateral: Characterized by planted-leg extension (extending the knee while the heel is planted) to push the system.
- Forward: Characterized by heel-first contact followed by knee flexion.
Efficiency vs. Tracking: While FS reduced CoT (improving energy efficiency), it increased tracking error. This indicates a trade-off where suppressing foot motion too aggressively hinders the propulsive motions necessary for velocity tracking.
Symmetry: SY successfully enforced symmetric leg behavior during longitudinal motion, whereas policies without SY often developed asymmetric strategies where one leg dominated propulsion.

C. Sim-to-Real Deployment

The policy was deployed on a physical Unitree G1. The robot successfully generated forward, backward, lateral, and turning motions while seated, using only proprioceptive inputs and velocity commands, demonstrating the viability of the approach without explicit contact sensing.

5. Significance and Claims

The paper claims that seated locomotion can be acquired by minimally extending standard standing velocity-tracking environments, provided that specific regularization and curriculum strategies are employed to handle the unique contact dynamics.

Complementarity: Seated behavior is presented as a complement to standing behavior, offering energy efficiency (no holding current for weight support) and the ability to change working positions without standing up.
Robustness: The study demonstrates that a policy trained without contact sensing or chair state estimation can generalize to a physical robot, suggesting that proprioception and velocity commands are sufficient for this complex contact-rich task.
Methodological Insight: The work highlights that while foot-slip penalties are beneficial for energy efficiency, they must be carefully balanced with curriculum learning or symmetry constraints to avoid catastrophic policy collapse (stationary local optima) in seated scenarios.

The authors do not claim superiority over standing locomotion in all metrics but emphasize that seated locomotion is a feasible, efficient, and distinct mode of operation that can be learned effectively using DRL with specific architectural and training adjustments.