Key takeaways
- The DDC paper first appeared on August 1, 2026, and was revised to v2 on August 14.
- The same center-of-mass position can be recoverable or unstable depending on its velocity.
- A support-foot frame removes the unmeasured base linear velocity from the relative balance signal.
- DDC achieved a clean single-leg hold on 89 of 90 held-out motions in a sim-to-sim benchmark.
- The physical Unitree G1 result is qualitative; quantitative hardware testing and transfer beyond G1 remain open.
Standing on one leg can be harder than performing a flashy motion
Researchers at Peking University posted version 1 of the DDC, or Deployable Dynamic-CoM, paper to arXiv on August 1, 2026 and revised it to version 2 on August 14. The starting observation is counterintuitive: recent humanoid controllers can track running, dancing and other large motions, yet often cannot hold a clean single-leg stance without putting the raised foot down or hopping on the support foot.
The authors do not count every non-fall as balance. A trial is Perfect only if the support foot never hops, the swing foot never touches down, the robot does not fall and body tracking remains inside the stated keypoint gate. A robot that stays upright by breaking the single-support condition is Marginal; a fall is Failure. This turns single-leg stance from a visually selected demo into a separately measured capability.
Why center-of-mass position alone reacts too late
Imagine balancing a pencil on a fingertip. Its center can still be over the fingertip while the pencil is already moving outward too quickly to save without moving the support. The same logic applies to a humanoid: position by itself does not reveal whether the center of mass is moving back toward the foot or accelerating away from it. Dynamic balance therefore needs both position and velocity.
The paper uses the capture point, or extrapolated center of mass, written as ξ=c+ċ/ω₀ with ω₀=√(g/h). In the simplified linear-inverted-pendulum picture, a capture point inside the support polygon can come to rest without changing the contact. If it leaves that region, ankle action alone may no longer be enough. This model is useful, but a many-joint humanoid with changing height and contacts is more complex than the equation.
A support-foot frame changes the sensing problem
Hardware creates a practical obstacle: an on-board sensor does not directly measure the pelvis or base’s absolute linear velocity. Estimators can add drift and delay, while external motion capture would defeat the goal of self-contained deployment. The authors instead express the center of mass relative to the support foot. The difficult base linear velocity enters both terms in the same way and cancels when their difference is taken.
The remaining state can be reconstructed from joint encoders, the pelvis IMU gyroscope, the kinematic and mass model, and a mask indicating which foot supports the body. The deployed actor receives four horizontal position-and-velocity balance terms inside a 463-dimensional observation and runs at 50 Hz. A privileged critic sees extra state during training, but it is discarded for deployment, so the same actor is used without teacher–student distillation.
The policy is rewarded for preventing imbalance, not merely surviving it
DDC’s rewards create both spatial and temporal buffers. A margin term reacts as the capture point approaches the support boundary, while a time-to-boundary term reacts when the predicted crossing time becomes short. The paper reports a lateral support half-width of roughly 2.75 cm in single support, showing why a static picture of the center over the foot can hide an imminent loss of balance.
The reward design also draws on human postural-control research: it encourages a distal-first response in which the ankle carries small corrections before the knee and hip become more active. A jerk penalty discourages high-frequency motor chatter that might survive in simulation but be undesirable on hardware. These are learned preferences, not formal safety guarantees or proof that the robot cannot fall.
The 89-versus-zero headline needs its denominator and rules
In a clean deterministic MuJoCo sim-to-sim evaluation, DDC produced a Perfect hold on 89 of 90 held-out motions, or 98.9%. Eight released general-purpose whole-body tracking policies achieved 0 of 90 Perfect holds under the same outcome definition. Many did not immediately fall: several survived by touching down the swing foot or hopping the support foot, which the benchmark deliberately labels recovery rather than genuine single-leg balance.
The denominator matters. The 90 test motions come from a stratified 900-motion set derived from 9,814 valid synthetic clips, and the single-support window is about 3.3 seconds. The baseline harnesses and physical plants are not identical for every method, so the paper treats cross-method continuous values as descriptive rather than a strict ranking. The 98.9% result must not be generalized to every floor, disturbance, payload or humanoid.
What the physical G1 demonstration actually establishes
The selected actor was deployed on a physical 29-DoF Unitree G1 through ONNX at 50 Hz. The paper and project page show the robot holding several distinct single-leg poses. That is meaningful evidence that the support-relative observation can be reconstructed on board and that the learned policy can run on real hardware. The released code, policy, data and benchmark also make independent reproduction more plausible.
However, the hardware section does not report a success rate comparable to 89 of 90, a force-versus-recovery envelope, repeated-duration statistics, varying-friction trials, or thermal and battery conditions. The authors explicitly describe the physical transfer as qualitative and list quantitative real-world metrics and deployment beyond the Unitree G1 as future work. Simulation competence and field reliability are not the same evidence level.
The product question is a test protocol, not a demo clip
For product evaluation, the right question is not simply whether a robot can stand on one leg. A useful protocol fixes support duration, swing-foot and pelvis height, floor friction, sensor noise, disturbance magnitude and direction, and whether re-contact is allowed. Separating clean hold, recovery and fall reveals more than a single survival number and makes comparisons harder to game with selected clips.
Commercial relevance appears in tasks that repeatedly create short single-support phases: narrow passages, stairs, uneven stepping stones or dressing assistance. Even there, DDC should be viewed as a balance observation and learning design, not a complete general-purpose robot stack. Product-level confidence will require repeated hardware trials, defined failure modes, and transfer across body shapes, foot sizes and payloads.
Sources reviewed
- arXiv: A Change of Frame Makes Balance Observable (v2, 2026-08-14)
- arXiv HTML: full paper and technical appendix
- DDC code and benchmark repository
- DDC stratified single-leg-balance motion set
Sources reviewed on September 22, 2026.
Related reading
Sean Woo — I have spent more than 15 years shaping robotics technology and business direction. Drawing on public technical documents, research papers and company announcements, I explain changes in robotics and AI in accessible terms. The interpretations in this publication do not represent the official position of any company or organization.