Source-linked AI summary
Natural Sit-to-Stand Motion Synthesis For Humanoids via Guided Assistance Curricula and Staged Rewards
Meet Pal Singh, Vyankatesh Ashtekar, Ashish Dutta
TL;DR
Humanoid sit-to-stand is challenging because many balanced motions are possible, especially from deep-seated postures. The paper learns such motion from scratch with a single proprioceptive PPO policy using IK pose randomisation, staged biomechanics-inspired rewards, and coupled assistance and chair-height curricula. In deterministic force-free evaluation, it achieves 97.8% balanced-standing success across eight chair lanes, while remaining limitations include occasional seat bracing and settling times up to 5 s.
Problem
Humanoid sit-to-stand requires smooth, controlled transitions from seated postures, but existing RL methods often use shallow initial postures or produce abrupt deep-chair movements.
Method
A single proprioceptive PPO policy learns from IK-randomised poses across eight chair heights using staged rewards and a coupled pelvis-assistance, action-bound, and chair-height curriculum.
Results
97.8% balanced-standing success is achieved in deterministic force-free evaluation across eight chair lanes.
Takeaways & Limitations
Rise-fraction staging and progressive chair unlocking are identified as key drivers of chair-height generalisation, while the learned motion is smooth and human-like.
Takeaways & Limitations
With 2–3 cm heel-to-seat clearance, the policy occasionally braces against the seat, and full settling can take up to 5 s.
Abstract
from arXiv · showhide
A humanoid has infinitely many ways to stand up from sitting while maintaining balance, making sit-to-stand (STS) a challenging control problem. We synthesise natural humanoid STS motion from scratch using reinforcement learning, without demonstrations or reference trajectories. A single Proximal Policy Optimisation policy learns smooth, human-like rising driven by three complementary components. (i) A coupled force/chair-height curriculum is used. A vertical pelvis-assist force aids early trajectory exploration and decays over training. Taller chairs are unlocked with decaying assisting force. This ensures that the policy masters a viable STS trajectory at each chair height before being exposed to harder ones, avoiding the premature distribution shift that otherwise collapses generalisation. (ii) Motion robustness is achieved by randomly sampling from a large number of inverse kinematics-generated initial and target poses spanning over eight chair heights. (iii) A set of rewards is defined inspired from biomechanics and optimal control studies. They shape the robot's angular momentum for seat-off, and enable support-region transition via centre of pressure attraction function to ensure smooth low-effort actuation. On a deterministic force-free evaluator, the policy attains more than 97% balanced-standing success across eight chair heights. The policy generalises smooth motion across chair heights and enables the robot to rise from substantially deep-seated postures as compared to the state of the art.
1 Introduction
The paper addresses smooth, natural humanoid sit-to-stand from deep-seated postures using pure reinforcement learning without demonstrations or reference trajectories. It combines physics-inspired rewards, IK-based pose randomisation, and coupled assistance and chair-height curricula.
- Humanoid sit-to-stand is essential for contact-rich whole-body transitions in human environments.
- Existing approaches span optimisation, reference-trajectory methods, pure RL, and RL with reference motion.
- Optimisation methods face contact discontinuities and require smooth objective gradients, while reference trajectories require robot-specific retargeting.
- State-of-the-art RL methods often use shallow seated postures or produce abrupt movements from deep chairs.
- The proposed method uses physics-based rewards, IK-randomised poses across eight chair heights, rise-fraction reward staging, and coupled force-decay chair unlocking.
- The resulting motion is smooth, balanced, and amenable to physical implementation.
2 Methodology
The method formulates sit-to-stand as a finite-horizon MDP solved with PPO, using proprioceptive control, physically consistent IK randomisation, staged rewards, and coupled curricula. Rewards regulate rise stages, contact support, angular momentum, bracing, posture, symmetry, smoothness, and effort.
- STS Formulation using RL: STS is formulated as a finite-horizon MDP and solved with PPO for a 29-DOF Unitree G1 in MuJoCo.The policy runs at 50 Hz, with episodes lasting up to 40 s or 2000 steps unless safety termination occurs.
- STS Formulation using RL: The proprioceptive policy observes a 97-dimensional state and outputs incremental joint commands mapped to PD targets and torques.The observation includes torso IMU quantities, joint positions and velocities, the previous action, and an action-bound multiplier.
- Randomisation, Rewards, and Curriculum: IK randomisation varies seat and stance heights, torso placement and pitch, foot positions, and arm degrees of freedom while preserving physically consistent poses.
- Randomisation, Rewards, and Curriculum: A rise fraction ρ=(hhead−h0)/max(hstand−h0,ϵ) provides chair-independent stage boundaries for low rise, mid rise, and standing.The boundaries are ρ<0.1, ρ<0.8, and ρ≥0.8, respectively.
- Randomisation, Rewards, and Curriculum: COP rewards attract support from the seat toward the feet during rising and toward the support-hull centroid while standing.The COP uses measured contact forces from nine sensors and does not require privileged balance signals.
- Randomisation, Rewards, and Curriculum: Angular-momentum, anti-parking, uprightness, foot-flatness, symmetry, standing-reference, smoothness, and power terms shape stable, low-effort motion.Forward pitch rate supports seat-off, standing terms damp it, and non-foot contacts are penalised.
- Randomisation, Rewards, and Curriculum: The coupled curriculum decays pelvis assistance and action bounds while progressively unlocking taller chair lanes after reliable performance.Table 1 defines the discrete seat-height lanes and their progressive unlocking as assistance decays.
3 Results and Discussion
The policy achieves high balanced-standing success across eight chair heights, while ablations clarify which curriculum and reward components support generalisation and smooth load transfer. The motion remains limited by occasional seat bracing and slower final settling.
- Overall performance: 97.8% balanced-standing success was achieved across eight chair lanes under deterministic, force-free evaluation.The evaluator used 50 resets per lane, for 400 success/fall trials.
- Ablations: Progressive chair unlocking and reward normalisation were necessary for chair-height generalisation.Sampling all chair lanes from the start reduced success, while absolute-height staging did not transfer across seats.
- Ablations: Removing sagittal pitch-momentum shaping lowered success, whereas removing COP shaping preserved headline success but more than doubled action jitter.The COP terms primarily regularised load transfer rather than determining headline success.
- Limitations: When heels began 2–3 cm from the seat edge, the policy occasionally braced against the seat, producing a less natural rise.The motion also slowed near completion: approximately 90% of the rise took 3–4 s, with full settling taking up to 5 s.
4 Conclusion
The work shows that pure reward-shaped RL can learn smooth, human-like STS from deep-seated postures across varied chair heights. Its single proprioceptive PPO policy combines IK pose diversity, staged rewards, COP and angular-momentum shaping, and coupled curriculum control.
- Conclusion: Pure reward-shaped RL learned smooth and human-like deep-seated humanoid STS motion.The conclusion frames this as achievable without relying on reference-motion tracking.
- Conclusion: An IK-generated pose library, rise-fraction staging, COP and angular-momentum shaping, and coupled force/action-bound curricula form the policy design.These components support training across multiple seat heights.
- Conclusion: More than 97% deterministic success was obtained across varied chair heights with one proprioceptive PPO policy.The conclusion identifies rise-fraction staging and progressive chair-height unlocking as key generalisation drivers.
- Conclusion: Future work will target sim-to-real validation.The reported conclusion establishes the current evidence primarily in deterministic evaluation.