A robot can know that it is leaning, know how fast it is moving, and still lack a feasible way to recover.

That is the next step after Week 15, From Pixels to State. Useful information supports a control decision. The body’s dynamics and its available actions determine what that decision can accomplish.

Balance belongs to the whole closed loop: the body, the feedback policy, the observations, the actuator and the environment. A statement about one part becomes misleading when it silently stands for all of them.

Posture leaves motion open

Imagine a body leaning five degrees forward. If it is moving farther forward, its recovery problem differs from the problem of a body already moving back toward upright.

Week 15 used this distinction to explain why perception must contain information relevant to action. Week 16 adds a further question: can the requested action actually change the motion in time?

For a transparent example, take an ideal pendulum with a fixed pivot. Its angle from upright is θ, its pivot inertia is I, and its gravitational torque coefficient is G. Delivered actuator torque is τ. With no external disturbance, the equation is

I θ̈ = G sin θ + τ.

Gravity initially increases a small forward lean. Feedback must counter that tendency while controlling motion. The pendulum is an explanatory model, with no foot slip, stepping or articulated body; its simplicity exposes assumptions rather than representing a complete humanoid.

Stability has assumptions

A proportional–derivative controller requests torque from angle and angular velocity:

τ = −kp θ − kd θ̇.

Under unlimited, immediate actuation and the small-angle approximation, the closed loop becomes

I θ̈ + kd θ̇ + (kp−G)θ = 0.

For positive inertia, this linear model converges to upright when kp>G and kd>0. The position feedback overcomes the destabilizing gravity term, and velocity feedback supplies damping.

These conditions establish a mathematical property of a specified model. They do not establish that an arbitrary push can be recovered, that a motor can provide the requested torque, or that a foot will remain in contact.

Local asymptotic stability means that sufficiently small deviations remain small and approach an equilibrium. A recovery claim must also state the initial conditions, permissible actions and goal. Lyapunov analysis provides tools for proving such scoped properties; a finite collection of successful trajectories supplies a different kind of evidence.1

Even a stable loop can leave error. In this simplified linear model, a constant external torque d produces an equilibrium angle d/(kp−G). Stability alone therefore does not mean perfect disturbance rejection.

Requested action has limits

Use hypothetical values I=1 kg m², G=9.81 N m, kp=20 N m/rad and kd=5 N m s/rad. Start at five degrees forward with angular velocity .4 rad/s forward. The requested initial torque is approximately −3.75 N m.

Our reproducible nonlinear simulation then changes only the torque cap or observation delay. With a 4 N m cap and no delay, the motion meets a declared settling criterion. With a 1 N m cap and no delay, it crosses an illustrative .5-radian angle boundary after about .634 seconds.

The second case has exact current state information. Its limitation is the action available to the body.

With the 4 N m cap, a .15-second observation delay meets the settling criterion in this example, while a .25-second delay crosses the boundary. Delay does not always cause failure; its consequences depend on the loop and initial conditions.

The calculation separates requested and delivered action. It reports model outcomes, not physical falls. The angle boundary and settling thresholds were chosen for explanation; they are not measured safety limits.

Contact changes the question

The fixed pivot hides another limitation: a real robot must obtain external forces through contact.

For a simple planar Coulomb friction model, a contact requires a nonnegative normal force and a tangential force no greater in magnitude than the friction coefficient times that normal force. Finite surfaces also constrain where pressure acts and which moments can be supported.2,3

A desired acceleration may therefore be incompatible with the assumed contact. Foot geometry contributes to balance, but geometry alone does not determine whether a moving body can return to a goal.

Stepping, body motion and additional contacts can change the available recovery strategies. Formal capturability research examines which states can reach a stopping condition under specified models and constraints; its conclusions depend on those choices.4

This is probably the most extreme in our current example. A humanoid can exploit strategies the model excludes. Failure of the pendulum controller cannot establish that the humanoid has no feasible recovery action.

Evidence must name its scope

Learning changes how a feedback policy is obtained; it does not remove the physical loop. Radosavovic and colleagues report physical humanoid locomotion and adaptation, while acknowledging falls under excessive disturbances.5 A 2026 RSS study by Wu and colleagues reports contact-rich parkour on physical hardware and identifies sensing and hardware limits.6

These results support specific capabilities. They do not settle every disturbance, surface, payload or repeated-operation condition.

When evaluating balance, I would ask for a record of the initial state, disturbance, observation timing, requested action, delivered action, contact assumptions and recovery endpoint. Failed attempts and the tested range matter alongside successful examples. The record is a proposed review practice; its added value remains unmeasured.

For education, the same distinction offers an assessment task. Ask learners to explain why a stable ideal calculation and a failed constrained simulation can both be correct. Require them to identify which assumption changed and which observation would test the explanation. This is a proposed activity, not evidence of improved learning.

The next step toward Week 17 is concrete. A hand on a support may supply new external forces and moments. Whether that contact helps depends on geometry, force, friction and timing. Grasp recognition alone cannot answer the question.

Balance analysis becomes more useful when it makes those conditions inspectable. The central question is not merely whether a controller can produce a correction. It is whether the coupled system can deliver a feasible correction before the declared recovery opportunity is lost.

Discussion invitation: What evidence would you require before extending a balance result to a different contact condition or disturbance regime?

References

1. Tedrake, R. (n.d.). Lyapunov analysis. In Underactuated robotics. Massachusetts Institute of Technology. Retrieved October 9, 2026.

2. Tedrake, R. (n.d.). Highly-articulated legged robots. In Underactuated robotics. Massachusetts Institute of Technology. Retrieved October 9, 2026.

3. Caron, S., Pham, Q.-C., & Nakamura, Y. (2015). Stability of surface contacts for humanoid robots: Closed-form formulae of the contact wrench cone for rectangular support areas [Preprint, version 1]. arXiv. The accessed manuscript supplies the cited theoretical and simulated findings.

4. Posa, M., Koolen, T., & Tedrake, R. (2017). Balancing and step recovery capturability via sums-of-squares optimization. Proceedings of Robotics: Science and Systems XIII.

5. Radosavovic, I., Xiao, T., Zhang, B., Darrell, T., Malik, J., & Sreenath, K. (2024). Real-world humanoid locomotion with reinforcement learning. Science Robotics, 9(89), eadi9579.

6. Wu, Z., Huang, X., Yang, L., Zhang, Y., Chen, X., Abbeel, P., Duan, R., Kanazawa, A., Sferrazza, C., Shi, G., & Liu, K. (2026). Perceptive humanoid parkour: Chaining dynamic human skills via motion matching. Proceedings of Robotics: Science and Systems XXII.

7. He, T., Gao, J., Xiao, W., Zhang, Y., Wang, Z., Wang, J., Luo, Z., He, G., Sobanbabu, N., Pan, C., Yi, Z., Qu, G., Kitani, K., Hodgins, J. K., Fan, L., Zhu, Y., Liu, C., & Shi, G. (2025). ASAP: Aligning simulation and real-world physics for learning agile humanoid whole-body skills. Proceedings of Robotics: Science and Systems XXI. Official abstract and bibliographic record only were assessed.

8. International Organization for Standardization. (2025). Robotics—Safety requirements—Part 1: Industrial robots (ISO Standard No. 10218-1:2025). Catalogue metadata and scope only.

9. International Organization for Standardization. (2014). Robots and robotic devices—Safety requirements for personal care robots (ISO Standard No. 13482:2014). Catalogue metadata and scope only.

10. International Organization for Standardization. (n.d.). Robotics—Safety requirements for service robots (ISO/FDIS 13482, final draft). Retrieved October 9, 2026. Catalogue status: under development; normative text not assessed.