Hey friend
If there is one thing that still holds humanoid robots back from being truly practical it is energy efficiency.
Todays best robots, like Tesla Optimus, Figure 01 and Unitree G1 can only walk for a relatively short time before their batteries run out. Humans can walk for hours on little energy. Why is it so hard for humanoid robots to do the same?
In this article we will explore the journey of making humanoid walking more energy-efficient from the idea of dynamic walking all the way to modern Model Predictive Control.
1. Passive Dynamic Walking: Learning from Nature
The most energy-efficient walkers in history were not programmed they were almost purely mechanical.
Humanoid robots like these can walk down a slope with zero actuation. They use gravity and the natural pendulum motion of the legs to keep moving.
Key principles of dynamic walking are:
- The legs swing like pendulums.
- Heel strike and energy loss are minimized through careful mechanical design.
- Very little energy is added during the gait cycle.
Robots like Cornell Ranger and Delfts Flame achieved efficiency using this approach. Some could walk with a Cost of Transport close to or even better than humans.
However passive walkers only work on slight downhill slopes. Cannot stop, turn or walk on flat ground. They taught us a lesson: the most efficient walking uses natural dynamics instead of fighting them.
2. From Passive to Semi-Walking
Modern humanoid designers try to borrow as much as possible from passive dynamics while adding just enough actuation to make the robot practical.
Techniques include:
- ankles and Series Elastic Actuators to store and release energy during the gait cycle.
- Underactuated design letting the legs swing freely during parts of the step instead of controlling every joint all the time.
- Ballistic walking phases, allowing the swing leg to move under gravity and momentum.
This semi-passive approach significantly reduces energy consumption compared to fully actuated walking.
3. Trajectory Optimization for Efficient Gaits
The next leap came from using optimization to find energy-efficient walking patterns.
Engineers define an optimization problem:
- Minimize energy
- While satisfying constraints, like stability and joint limits
This is usually solved using trajectory optimization methods. The result is a efficient gait that exploits natural dynamics as much as possible.
4. Model Predictive Control for Real-Time Efficiency
While trajectory optimization gives offline gaits real robots need to react to disturbances.
Model Predictive Control takes optimization online:
- At every time step it solves a short-horizon optimization problem.
- It predicts the robots motion and chooses the best actions that minimize energy while keeping the robot balanced.
- It can continuously adapt to changes in terrain, pushes or payload.
Many of the efficient walking controllers today combine offline trajectory optimization with online Model Predictive Control.
Current Best Practices for Energy-Efficient Walking
From what we see in research and industry like Tesla Optimus and Atlas:
- Use compliant actuators in the legs
- Design trajectories that exploit natural pendulum dynamics
- Include energy terms in the cost function of Model Predictive Control and trajectory optimization
- Use whole-body coordination like arm swing
- Apply imitation learning from efficient human walking data
- Heavily use simulation to find robust low-energy gaits
Some recent research humanoids have achieved Cost of Transport values approaching 0.3–0.5 getting closer to human efficiency.
My Personal Take
Energy efficiency is not about using better motors or lighter materials. It is fundamentally about working with physics of against it.
The promising path forward is a hybrid approach:
- Passive dynamic principles for baseline efficiency
- Optimal control and Model Predictive Control for real-time adaptation and stability
- Machine learning to discover clever energy-saving behaviors
I believe the humanoid robots that will eventually walk for hours on a charge will not be the ones with the strongest motors but the ones that move most intelligently recycling energy using gravity wisely and minimizing unnecessary effort.
We have come a way from the stiff power-hungry walkers of the early 2000s. The combination of mechanical design, compliant actuators and smart control is finally starting to close the efficiency gap, with humans.