Most headlines about humanoid robots are about the "brain": large models that understand a scene and an instruction. Far fewer are about the layer underneath, the controller that keeps a humanoid upright while it walks, lifts, leans and gets bumped. In October 2026 that layer had a remarkable month on arXiv. Researchers published new methods for humanoid whole-body control that train a motion tracker on a single GPU in hours, carry loads against the chest, catch boxes thrown from five metres away and steer how a fallen robot stands back up. Almost all of them were validated on the same commercially available robot, the Unitree G1.
In this article I explain what whole-body control is, summarise what five of these papers actually report, and share how I read this kind of research as a technology executive.
What humanoid whole-body control is, and why it is the bottleneck
A humanoid has dozens of joints that must move together. Walking, reaching and carrying are not separate problems: when the arms hold a weight, the robot's centre of mass moves and the legs must compensate. A whole-body controller (WBC) is the policy that turns a compact command (walk at this speed, hold the torso at this height, put the hands here) into joint actions that keep the robot balanced.
The HULK paper describes the architecture the field has converged on: a learned WBC sits at the bottom and planners, teleoperation interfaces or agentic pipelines sit above it, issuing commands "without reasoning about balance themselves" and treating the WBC as a low-level "System 0". The authors note that massively parallel reinforcement learning in simulation has made this layer reliable enough to serve as that foundation, with rapid progress in agility and perception, while forceful interaction has remained comparatively underexplored.
For anyone evaluating humanoid robots, this matters. A robot that cannot keep its balance under real loads and surprises cannot do useful work, however capable its language model is.
Five October 2026 papers on humanoid whole-body control
1. SuperMimic: motion tracking on one GPU instead of 128
The dominant way to teach a humanoid expressive, general movement is to have it track large datasets of human motion. According to SuperMimic, from authors including Pieter Abbeel and Koushil Sreenath (accepted as an oral at CoRL 2026), this has been expensive: pipelines either merge per-motion specialists through lossy distillation or brute-force one policy with cluster-scale compute.
SuperMimic instead learns one differentiable world model of the motions' shared dynamics and trains it together with the tracking policy, using analytic gradients through short imagined rollouts. The reported results are striking: in 3 hours on a single GPU it surpasses the return of a state-of-the-art baseline trained for 7 days on 128 GPUs, cutting GPU-hours by more than 1,000 times. It fails on only 25 of 129,514 benchmark motions versus 179 for the baseline, and reaches zero failures after 9.8 hours on two GPUs. The policy transfers to a physical Unitree G1 across diverse and acrobatic motions.
2. HULK: carrying heavy loads with the whole body
HULK, from the University of Michigan, UCLA, Caltech and Amazon, tackles what the authors call forceful loco-manipulation: lifting, pushing and carrying heavy objects, including against the torso. It uses model predictive control (MPC) to guide reinforcement learning, trains two teachers (one resisting forces at the wrists, one walking while holding large objects against the body), adds a capture-point control barrier function during training to improve balance, and distils both into a single policy.
On a real Unitree G1, the distilled policy completed 5 of 6 cart push/pull trials at a common 150 kg load, against 3 of 6 for the best baseline, and pulled up to 300 kg in a separate sweep. It completed 4 of 4 trials carrying a mini-fridge in front of the torso and a canister at its side, where the baselines completed at most 2 of 4 and 0 of 4. In simulation the wrist-force teacher withstood torso pushes of up to 130 N. The authors are explicit about limits: grasp acquisition, regrasping and rapid contact changes remain outside the demonstrated scope, and the training-time barrier provides no runtime safety guarantee.
3. Reflex: reacting to a thrown box from egocentric vision
Reflex, from the University of Washington, Stanford and collaborators, studies a deceptively human skill: catching a thrown box with the whole body, using only the robot's own RGB-D camera. It learns in three stages: whole-body catching from state, then inferring a box's dynamics from delayed and incomplete observations, then recovering that representation directly from vision while the controller stays fixed.
In simulation Reflex reaches 85.7% catch success from egocentric RGB-D, close to 87.9% with privileged box data; removing temporal visual history cuts success by 35.6 and 39.8 points. On a real Unitree G1 it catches human-thrown boxes 65% of the time at 2.5 m and 40% at 5 m, which the authors put at 76% and 59% of human catchers' success at the same distances.
4. RISE: choosing how to get up after a fall
Falls will happen in real deployments. RISE, from the University of Michigan, starts from the observation that the same fallen pose admits several ways to stand, with different routes and final headings. It learns a vocabulary of recovery movements and a visual diffusion planner whose predictions can be steered at inference time toward a desired heading, rise speed or path, without retraining. The authors show diverse recoveries and terrain adaptation in simulation, complemented by real-world demonstrations on a Unitree G1.
5. SP-HLM: a cleaner interface between "what" and "how"
SP-HLM, from Shanghai Jiao Tong University, Moore Threads and the University of Hong Kong, focuses on the interface between a high-level policy and whole-body control. Instead of commanding every joint, the high-level model specifies just five task-critical poses (pelvis, both wrists, both ankles); a fast motion generator fills in coordinated full-body motion and a tracker executes it. The authors report a 52.75% overall success rate across eight simulated loco-manipulation tasks, ahead of the token-space and joint-space baselines they tested, and real-world success rates of 13/20 for object placement and 5/20 for cabinet opening on a Unitree G1. Their failure analysis is instructive: tasks that demand precise hand or foot placement while balancing on one leg were the hardest.
What these results tell a technology leader
Reading these papers together, I see four signals.
- Training cost is collapsing. If SuperMimic's results hold up, robust motion tracking stops being a cluster-scale project. Cheaper experimentation means faster iteration for labs and, eventually, for vendors adapting robots to specific tasks.
- A shared, affordable platform is accelerating research. Every paper above reports results on a Unitree G1. A common robot makes results easier to compare and reproduce, the same way shared benchmarks accelerated machine learning.
- The frontier is moving from motion to physical work. Carrying 7 kg against the torso, pulling a 300 kg cart and catching a box are steps toward the logistics and manufacturing tasks the HULK authors name explicitly. They are still lab results with small trial counts.
- Honest numbers matter more than demo videos. A 40% catch rate at 5 m or 5/20 on cabinet opening is real progress and also a clear statement that reliability is not yet industrial. The best papers publish their failure cases; buyers should ask vendors for the same.
How to evaluate humanoid robot claims
When a vendor or internal team pitches a humanoid pilot, I use the research above as a checklist:
- Ask which layer is being shown. Is the demo about the high-level model, the whole-body controller, or teleoperation? Each fails differently.
- Ask for success rates, not clips. How many trials, under what load, and what counts as failure (a fall, a dropped object, a protective stop)?
- Test disturbances and loads. Push recovery, payload limits and behaviour on uneven floors say more about readiness than a smooth walk.
- Ask about falls and recovery. How does the robot fall, how does it get up, and can operators control that behaviour?
- Separate simulation from hardware. Many headline numbers are simulation results; hardware trials are usually fewer and harder.
- Plan for supervision. Even the strongest October results leave room for failure, so pilots need human oversight, safety zones and clear stop procedures.
In Physical AI in 2026 I looked at the vision-language-action models that act as the robot's "brain"; this article covers the layer that keeps that brain's body upright. And as with AI agents in practice, the useful question is not whether the system can perform once, but whether it can do so reliably, observably and safely.
Closing thought
October 2026 did not produce a humanoid that is ready for every warehouse. It produced something arguably more important: evidence that the foundation layer of humanoid robots is getting cheaper to train, stronger under load and better at reacting to the unexpected, with results published openly enough to check. For leaders, the right posture is informed patience: track this research, insist on measured reliability, and design pilots around the tasks where today's numbers are already good enough.
References
- SuperMimic: Scaling Humanoid Whole-Body Control with One Differentiable World Model, Q. Liao, Y. Shao, J. Long, K. Zakka, R. Horowitz, P. Abbeel, K. Sreenath, arXiv:2610.28480, 20 October 2026
- HULK: Learning Whole-Body Forceful Loco-Manipulation for Humanoids, A. Dang et al. (University of Michigan, UCLA, Caltech, Amazon), arXiv:2610.08970, October 2026
- Reflex: Visually Grounded Reactive Humanoid Control, T. Jia, W. Huang et al. (University of Washington, Stanford and others), arXiv:2610.12772, October 2026
- Test-Time Motion Steering for Perceptive Humanoid Fall Recovery, O. Azulay, C.-L. Hsieh, S. X. Yu (University of Michigan), arXiv:2610.30061, October 2026
- SP-HLM: Sparse-Pose-Driven Motion Generation and Whole-Body Control for Humanoid Loco-Manipulation, Z. Yang, Q. Lin, C. Shi et al. (Shanghai Jiao Tong University, Moore Threads, University of Hong Kong), arXiv:2610.30044, October 2026