How Robot Toes Cut Energy Use by 17.5% and Redefined Agility
17.5%. That’s the drop in energy cost of transport (CoT) for a bipedal robot simply by adding active toes—mechanical joints that mimic the push-off and shock absorption of human feet. In a world where every watt counts for mobile robots, that’s not incremental progress. It’s a leap. And it comes not from a new battery or a lighter material, but from a design detail long overlooked: the toe.
This isn’t just about efficiency. The same robotic feet reduced heel-strike ground reaction forces by 5.0%—a measure of impact that, in humans, correlates with joint wear and injury risk. In agility tests, they cut average path deviation by 25.0% and maximum deviation by 34.0%, meaning robots could navigate tight, dynamic environments with far greater precision. These numbers, drawn from high-fidelity simulations by Kim et al. (2026), suggest that the future of humanoid robotics may hinge on getting the foot right.
For decades, humanoid robots have walked with flat, rigid feet or simple hinges—functional, but far from optimal. The human foot, by contrast, is a marvel of biomechanical engineering: 26 bones, 33 joints, and over 100 muscles and tendons, all working in concert to absorb shock, store elastic energy, and propel us forward. The toes alone contribute up to 40% of push-off force during walking (Neumann, 2016). Yet most robots have treated the foot as an afterthought.
Now, with advances in actuation, simulation, and machine learning, researchers are finally able to test what was long suspected: that active toes aren’t just biomimetic flourishes—they’re performance multipliers.
The Science
The study introduces a 14-degree-of-freedom (DOF) bipedal robot designed to emulate human-like leg dynamics, with a key innovation: active toes. Each foot features a 2-DOF ankle and a 1-DOF toe joint, mirroring the complexity of human foot mechanics. The robot’s legs weigh 8.27 kg each, with all heavy actuators housed in the thigh to minimize distal mass—a design choice that enhances agility and mimics human weight distribution (
).
But hardware alone isn’t enough. To isolate the impact of active toes, the researchers built a high-fidelity simulation environment using Isaac Lab (Mittal et al., 2025), a GPU-accelerated platform that models real-world actuator dynamics, friction, and power consumption. Crucially, this simulation accounts for cooperative actuation—a design where the knee, ankle, and toe share motors through a belt-and-linkage system. This coupling allows multiple joints to contribute to propulsion, much like human muscles span multiple joints.
The Jacobian matrix maps motor velocities to joint velocities ($\dot{\theta} = J_m \dot{\phi}$), and its transpose links motor torques to joint torques ($\tau_m = J_m^T \tau_{jt}$). This means a single motor can drive knee extension, ankle plantar flexion, and toe push-off simultaneously—amplifying force during toe-off.
To train the robot, the team used proximal policy optimization (PPO), a reinforcement learning (RL) algorithm. They trained two versions of the robot: one with active toes (14-DOF), and one with the toe joint removed (12-DOF), otherwise identical. Both used the same observation space, reward function, and training procedure—ensuring a fair comparison.
The reward function was minimal, focusing on task completion and energy use. But instead of the common proxy of penalizing squared joint torques, the researchers directly minimized CoT, defined as:
where is resistive heating, is mechanical power, and , , are robot mass, gravity, and velocity. This approach captures real-world energy losses, including motor inefficiencies and regenerative braking.
To prevent the policy from exploiting unrealistic motor behaviors, the team added a soft thermal constraint inspired by recent work on quadruped robots (Qian et al., 2026). This penalty discourages sustained high-torque states that could overheat motors, nudging the robot toward sustainable, efficient gaits.
What They Found
At a walking speed of 1.33 m/s—roughly 3 mph, a typical human pace—the toe-equipped robot consumed 17.5% less energy than its toe-less counterpart. Total power dropped from 313.0 W to 260.2 W, with the largest savings in Joule heating (−24.9%) and mechanical losses (−7.6%)
Total Power Consumption
Total power consumption during straight-line walking at 1.33 m/s
| Label | Value |
|---|---|
| Toe-equipped | 260.2 |
| Toe-ablation | 313 |
.
Total Power Consumption
Total power consumption during straight-line walking at 1.33 m/s
| Label | Value |
|---|---|
| Toe-equipped | 260.2 |
| Toe-ablation | 313 |
This efficiency gain wasn’t free. The toe-equipped robot had to generate more force at toe-off, but it did so in a way that reduced load on other joints. A breakdown of joint-level power (
) shows that without active toes, the hip, knee, and ankle had to compensate: hip power increased by 22%, knee by 11%, and ankle by 29%. The toe wasn’t just helping—it was redistributing work, allowing the robot to walk with less strain on its larger, more power-hungry joints.
Impact absorption also improved. Heel-strike ground reaction force (GRF)—the jolt felt when the foot first hits the ground—dropped from 1021.1 N to 970.0 N, a 5.0% reduction. While this may seem modest, in robotics, even small reductions in peak forces can extend hardware lifespan and improve stability on uneven terrain. In humans, such forces can reach 2.5 times body weight during running (NASA, 1995); for a 32 kg robot, that’s over 780 N. Reducing impact isn’t just about efficiency—it’s about durability.
In the agility test—a robotic version of the athletic T-test—the toe-equipped robot didn’t move faster. In fact, its average speed was slightly lower (2.085 m/s vs. 2.182 m/s). But it followed the path more precisely. Average path deviation shrank from 0.308 m to 0.231 m (−25.0%), and maximum deviation from 1.033 m to 0.682 m (−34.0%)
Agility Test: Path Deviation
Path deviation during T-test agility trial
| Label | Value |
|---|---|
| Toe-equipped | 0.231 |
| Toe-ablation | 0.308 |
| Toe-equipped | 0.682 |
.
Agility Test: Path Deviation
Path deviation during T-test agility trial
| Label | Value |
|---|---|
| Toe-equipped | 0.231 |
| Toe-ablation | 0.308 |
| Toe-equipped | 0.682 |
This suggests a fundamental shift: the robot wasn’t just moving—it was maneuvering. On sharp turns, the active toes provided better grip and controlled push-off, reducing skidding and drift. The toe-ablation robot, lacking this fine control, overshot corners and wobbled through transitions (
).
These results weren’t artifacts of simulation. The researchers went to great lengths to minimize the sim-to-real gap: modeling motor torque-speed curves, transmission losses, and friction. The actuators used in simulation—U10 cycloidal gears for hips, BLDC motors with planetary gears for knees and ankles—are real, off-the-shelf components. The control policy maps joint commands to motor torques via the CA Jacobian, then clips them against 4-quadrant torque-speed envelopes, ensuring feasibility.
Why This Changes Things
For years, humanoid robotics has been caught in a paradox: we build machines to operate in human environments, yet their movements are anything but human. They walk stiffly, avoid stairs, and struggle on gravel. Much of this comes down to feet.
Consider Boston Dynamics’ Atlas or Tesla’s Optimus: both use flat, rigid soles. They’re stable, but inefficient. The average human CoT is around 0.316; most humanoid robots exceed 1.0 (Kim and Wensing, 2017). That means they burn more than three times the energy per unit weight and distance. Such inefficiency limits battery life, increases heat, and restricts real-world use.
This study shows that a 17.5% reduction in CoT is possible without new materials or larger batteries—just by adding a single actuated joint. That’s equivalent to extending a robot’s operational time by over 20%, or reducing its battery size (and cost) for the same mission duration.
But the implications go beyond energy. The 5.0% drop in heel-strike GRF suggests that active toes could make robots safer for human interaction. Lower impact forces mean less vibration, less noise, and reduced risk of damage to floors—or to the robot itself. In industrial settings, where robots may walk thousands of kilometers, this could translate to fewer maintenance cycles and longer service life.
The agility gains are perhaps most striking. A 34.0% reduction in maximum path deviation isn’t just a number—it’s the difference between navigating a cluttered warehouse and colliding with a pallet. In search-and-rescue scenarios, it could mean the difference between reaching a survivor and getting stuck in rubble.
And this is just the beginning. The current design uses a single-DOF toe, approximating the metatarsophalangeal joint. Human toes have multiple degrees of freedom and intrinsic muscles that allow fine adjustments on uneven ground. Future robots could incorporate multi-DOF toes, soft materials, or even artificial tendons—each a potential source of further gains.
The study also challenges a common assumption in robotics: that simplicity is always better. Many engineers avoid complex feet because they fear fragility or control difficulty. But here, complexity—when grounded in biomechanics and validated with rigorous simulation—pays off. The toe didn’t make control harder; it made the whole system more efficient.
What’s Next
The most immediate next step is hardware deployment. The robot exists in simulation, but not yet in physical form. Building and testing it will be the true test of whether these gains survive the transition to reality. Sim-to-real gaps persist, especially in contact-rich tasks like walking. Friction, wear, and sensor noise could erode some of the benefits.
Still, the researchers are optimistic. By modeling actuators and transmissions in detail, they’ve built a bridge between simulation and reality. The use of real motor specs, friction models, and power electronics suggests the simulation isn’t just illustrative—it’s predictive.
Future work could explore terrain adaptability. The current tests were on flat ground. How would active toes perform on gravel, grass, or stairs? Would they enable running, jumping, or dancing? The cooperative actuation design—where knee, ankle, and toe work together—hints at potential for dynamic motions beyond walking.
Another open question is scalability. Could this design work on smaller or larger robots? On quadrupeds? The principles of distal actuation and impact modulation may apply broadly. Indeed, some quadruped robots already use toe-like structures for better grip (e.g., MIT’s Mini Cheetah). But bipeds, with their narrow support base, may benefit even more from precise foot control.
There’s also a design trade-off to consider: added complexity. The active toe adds weight, cost, and potential failure points. Is a 17.5% efficiency gain worth it? For some applications—delivery robots, factory workers, elder-care assistants—the answer may be yes. For others, simplicity may still win.
Finally, this work opens a broader conversation about biomimicry in robotics. We’ve long copied birds to build planes, fish to design submarines. Now, we’re finally starting to copy the human foot—not just in shape, but in function. The toe, once dismissed as a vestigial appendage, may turn out to be one of the most important innovations in mobile robotics.
As Kim et al. write: “Although these simulation-first insights await physical hardware validation… they provide a rigorous foundation for closing the sim-to-real gap.” That foundation isn’t just about toes. It’s about a new way of building robots—one that respects the elegance of biology, leverages the precision of simulation, and aims not just to move, but to move well.