## Can a Humanoid Learn to Stand Up Without Ever Watching a Human Do It?

Yes — and new research from Meet Pal Singh, Vyankatesh Ashtekar, and Ashish Dutta demonstrates exactly that. Their paper, posted to arXiv on August 24, 2026, trains a single Proximal Policy Optimisation (PPO) policy to execute smooth, human-like sit-to-stand (STS) motion from scratch — no motion capture, no demonstrations, no reference trajectories. On a deterministic, force-free evaluator, the policy achieves more than 97% balanced-standing success across eight distinct chair heights, and generalises to substantially deep-seated postures beyond the current state of the art.

This matters because STS is deceptively hard. A humanoid has, in theory, infinitely many ways to rise from a chair while maintaining balance — the policy must not only reach a standing configuration but do so with low-effort, smooth actuation that won't stress actuators or destabilise the platform mid-rise. Solving it without demonstrations removes the data-collection bottleneck that plagues [imitation learning](https://humanoidintel.ai/glossary/imitation-learning) pipelines and opens a path to scalable [whole-body control](https://humanoidintel.ai/glossary/whole-body-control) of transitional motions.

---

## Three Engineering Choices That Make This Work

The architecture is a single PPO policy, but the training scaffolding is where the real engineering lives. The authors identify three complementary components:

### 1. Coupled Force/Chair-Height Curriculum

A vertical pelvis-assist force supports the robot during early training, giving the policy a viable trajectory to explore before it has learned anything useful. Critically, this force decays over training — and taller chairs are only unlocked as the assist force diminishes. The sequencing is deliberate: the policy must master STS at a given chair height before being exposed to harder configurations. The authors explicitly flag the problem this solves — "premature distribution shift that otherwise collapses generalisation." This is a known failure mode in curriculum RL for contact-rich tasks, and the coupled decay mechanism is a clean solution.

### 2. Pose Diversity via Inverse Kinematics

Motion robustness comes from sampling across a large number of initial and target poses generated by [inverse kinematics](https://humanoidintel.ai/glossary/inverse-kinematics), spanning the full eight-chair-height range. This is essentially domain randomisation applied to the configuration space rather than physical parameters — forcing the policy to learn a generalised rising strategy rather than overfitting to a narrow set of start conditions.

### 3. Biomechanics-Inspired Reward Shaping

The reward structure draws on biomechanics and optimal control literature. Two components stand out analytically:

- **Angular momentum shaping at seat-off** — the moment of leaving the chair is the most mechanically challenging phase; shaping angular momentum here directly addresses the instability window.
- **Centre of pressure attraction function** — this steers the robot's support-region transition, ensuring the centre of pressure moves smoothly as weight shifts from seat to feet. The result is low-effort actuation through what is otherwise a highly nonlinear contact transition.

Together, these reward terms encode what physical therapists and biomechanists already know about efficient human rising — then express it as a differentiable signal the RL agent can optimise.

---

## Why the Demo-Free Framing Is Significant

The field has leaned heavily on motion capture and teleoperation data to generate reference trajectories for [sim-to-real transfer](https://humanoidintel.ai/glossary/sim-to-real-transfer). That works well for locomotion gaits where human data is abundant, but transitional motions — sit-to-stand, getting up from the floor, climbing stairs from a seated position — are harder to capture cleanly and vary enormously across body geometries and furniture configurations. A policy that synthesises natural motion from reward signals alone sidesteps the data-collection pipeline entirely.

The authors' claim of outperforming the state of the art specifically on deep-seated postures is worth scrutinising. Deep-seated postures are the worst case for STS: the hip angle is more acute, the contribution of hip extensors is reduced, and the centre of mass starts further from the support polygon. If the policy genuinely handles these cases at 97%+ success without assist forces at evaluation time, that's a meaningful advance over approaches that either require near-standard chair heights or depend on residual assist forces to close the gap.

---

## Industry Implications: Who Should Care

Every humanoid developer deploying into real environments — factories, hospitals, elder care, logistics — will eventually face the STS problem. It's not a niche capability. Workers sit down. Robots that assist workers, or work alongside them, need to manage seated-to-standing transitions for themselves and ideally help humans do the same. [Agility Robotics](https://humanoidintel.ai/companies/agility-robotics), whose Digit platform is already in warehouse deployments, and [Figure AI](https://humanoidintel.ai/companies/figure-ai) targeting automotive and logistics environments, both operate in contexts where constrained-space transitions are daily requirements.

The broader research implication is methodological. If curriculum RL with biomechanics-informed rewards can solve STS from scratch, the same architectural pattern — graduated difficulty, domain-randomised initial conditions, physically-grounded reward terms — is a strong candidate for other transitional motion problems: floor recovery, stair negotiation, ladder climbing. The specific mechanisms here (coupled force/height decay, CoP attraction) may transfer more broadly than the STS application suggests.

One important caveat: the paper reports simulation results. [Zero-shot generalization](https://humanoidintel.ai/glossary/zero-shot-generalization) from sim to real for contact-rich tasks like STS remains genuinely difficult — furniture compliance, floor friction variation, and actuator dynamics all introduce gaps that simulation may not capture. Whether this policy survives hardware deployment without significant fine-tuning is the open question the arXiv version cannot answer.

---

## Key Takeaways

- A PPO policy trained by Singh, Ashtekar, and Dutta achieves **more than 97% balanced-standing success** across eight chair heights with no demonstrations or reference trajectories
- Three components drive performance: a **coupled force/chair-height curriculum**, **IK-generated pose diversity** across eight chair heights, and **biomechanics-inspired reward shaping** including angular momentum and centre of pressure terms
- The curriculum's key innovation is preventing premature distribution shift by coupling assist-force decay to chair-height unlocking
- The policy outperforms the state of the art on deep-seated postures — the hardest STS configuration
- All results are in simulation; sim-to-real transfer for contact-rich STS remains an open validation gap
- The methodology — graduated curriculum, domain-randomised configurations, physics-grounded rewards — is potentially applicable to other transitional motions beyond STS

---

## Frequently Asked Questions

**What is sit-to-stand (STS) and why is it hard for humanoid robots?**
STS is the transition from a seated to a standing posture. It's mechanically challenging because there are infinitely many feasible trajectories, the centre of mass must shift from a large support base (seat + feet) to feet alone, and the motion involves nonlinear contact transitions — all while maintaining balance. For robots, this is compounded by actuator limits and the difficulty of simulating chair-contact dynamics accurately.

**What is a PPO policy in the context of humanoid robot control?**
Proximal Policy Optimisation (PPO) is a reinforcement learning algorithm that trains a neural network policy by iteratively collecting environment interactions and updating the policy within a constrained step size to prevent instability. In humanoid control, a single PPO policy typically maps proprioceptive state observations to joint torques or position targets.

**How does the curriculum work in this paper?**
A vertical pelvis-assist force helps the robot explore viable STS trajectories early in training. As training progresses, the assist force decays. Harder chair heights (lower chairs requiring more work) are only introduced as the assist force decreases — ensuring the policy is competent at each difficulty level before advancing.

**Does this policy work on real hardware?**
The paper reports simulation results. Real-hardware validation is not described in the arXiv preprint. Sim-to-real transfer for contact-rich transitional motions is an active research challenge; gap characterisation would require physical deployment experiments.

**What makes the reward design different from standard locomotion rewards?**
Standard locomotion rewards primarily shape velocity, stability, and energy. This paper adds biomechanics-specific terms: angular momentum shaping at the seat-off moment (the highest-risk instability window) and a centre of pressure attraction function to manage the support-region transition smoothly. These terms encode domain knowledge from physical therapy and optimal control research rather than generic control heuristics.