A robot body designed with its controller
Humanoid robotics usually divides a difficult problem into two jobs. Mechanical engineers choose the body, joints and motors. Control researchers then teach that finished machine to walk, balance or copy a person. A new research platform called BRIDGE asks whether that order leaves performance on the table. Its designers use controller performance to help decide what the body should be.
The work appeared as a first-version arXiv preprint on September 3, 2026. Jianren Wang and colleagues from Carnegie Mellon University, Huazhong University of Science and Technology and JoyIn AI describe an 88-centimeter humanoid with 21 powered degrees of freedom. The project page reports a mass of 12.5 kilograms and an estimated platform cost of about $1,500.
The important contribution is not another compact robot performing an athletic demonstration. It is a development loop that connects human motion data, mechanical constraints and learned control before the design is frozen. If that loop holds up under independent reproduction, research teams could build bodies that are easier to control instead of spending more software effort compensating for poorly placed joints or unsuitable motors.
How the co-design loop works
The researchers begin with SMPL, a digital model commonly used to represent human body shape and movement. Human joints can rotate in ways that are difficult to reproduce with compact rotary actuators. BRIDGE therefore does not simply shrink a human skeleton. The team tests which robotic joint arrangements preserve useful human motion while remaining physically buildable.
One example is the waist. The researchers first compared two-axis combinations using motion-retargeting error, a measure of how much a human movement changes when mapped onto a robot. They then compared one-axis waist designs through closed-loop control and selected yaw rotation. This reduced the machine to 21 active degrees of freedom while keeping the joint structure useful for the target motion set.
Motor choice also enters the loop. The model includes actuator size, mass, inertia and calibrated torque-versus-speed limits. When a simulated movement fails because a joint cannot supply enough torque or speed, the process can assign a stronger actuator and reconstruct the surrounding links. The final design uses lighter 10-newton-meter-class actuators in the upper limbs, 55-newton-meter-class units at the hip-pitch and knee-pitch joints, and 25-newton-meter-class units elsewhere. In plain terms, the robot spends weight and power where demanding movements need them most.
What was measured in simulation
The paper compares BRIDGE with Bumi, Booster K1 and ToddlerBot under a common simulated control setup. The researchers trained SONIC motion-tracking policies for each morphology and evaluated them in MuJoCo using a benchmark assembled from the LaFAN1 and bones_seed motion collections. This is a controlled software comparison, not a race among four physical robots.
BRIDGE recorded a 94.83 percent simulated motion success rate. K1 reached 92.66 percent, Bumi 91.87 percent and ToddlerBot 88.23 percent. BRIDGE also had the lowest reported errors for joint position, joint velocity, root velocity and keypoint position. When the motions were separated into balance, highly dynamic and daily categories, its reported success rates were 95.00, 94.50 and 94.99 percent.
The authors also propose a human-likeness score that gives equal weight to motion-retargeting error and dynamic tracking error. On that metric, BRIDGE scored 0.5252, ahead of Bumi at 0.4321, K1 at 0.4198 and ToddlerBot at 0.3883. Those numbers support the narrower conclusion that BRIDGE performed best within the authors' chosen models, datasets, controllers and metric. They do not establish universal superiority over other humanoids or predict reliability in homes, factories or public spaces.
What the physical robot demonstrated
The researchers did build BRIDGE and show it moving outside simulation. The paper and project page present videos of locomotion and whole-body teleoperation, along with single-leg balance, bending, lunging, a Charleston dance sequence and a backflip. These demonstrations matter because a design optimized in software can fail when motors, gearboxes, structural flex and imperfect calibration enter the system.
They are still qualitative demonstrations. The paper does not report a physical head-to-head trial in which BRIDGE and the three comparison platforms perform the same motion protocol under matched conditions. It also does not provide long-duration reliability, fall-recovery, battery-life, payload or manipulation benchmarks. The simulation table and the hardware videos answer different questions and should not be merged into one performance claim.
The small scale is both a practical advantage and a boundary. A lighter machine can reduce laboratory cost and make falls less destructive. The authors also acknowledge that BRIDGE has limited workspace and payload capacity for complex manipulation. Its rotary motors produce packaging compromises that tendon-driven systems might reduce. The present evidence supports a research platform for motion and control, not a general-purpose labor robot.
The $1,500 estimate is not a deployment price
The paper's comparison table lists BRIDGE at $1,500, and the project page describes approximately the same figure as the platform cost. That is a notable target for humanoid research hardware, but neither source provides a complete cost accounting at the time of verification. The figure should be treated as the researchers' estimate for the platform, not the price of a ready-to-use deployment.
A complete build can involve fabrication, shipping, taxes, batteries, onboard or external computing, wiring, tools, assembly labor, calibration, spare actuators and damaged parts. A laboratory may also need motion-capture or teleoperation equipment, protected test space and staff time for safe operation. The current public materials do not specify which of those items are inside the estimate.
This distinction matters because low component cost can genuinely broaden access while still leaving a substantial integration bill. BRIDGE could lower the hardware barrier for university and independent robotics teams. Confirming that benefit will require a reproducible bill of materials, documented build time and reports from groups other than the original developers.
Open-source promise versus available artifacts
The paper repeatedly describes BRIDGE as open-source and says the platform is released with its control policy. The live project page is more cautious. It currently links a STEP-format CAD model, while both the code and tutorial are labeled as coming soon. The page says the complete assembly tutorial, electrical specifications, bill of materials, and training and deployment code will be released after the paper is accepted.
That gap does not erase the technical work, but it limits what outsiders can verify today. A CAD file can expose geometry. It cannot by itself reproduce the electronics, calibrated actuator limits, training pipeline, controller or safety procedures. Until the promised materials arrive, the stronger description is that BRIDGE has begun an open release rather than that a complete reproducible stack is already available.
The next meaningful evidence will come from public code, exact parts and build instructions, independent replicas and physical tests that measure endurance, energy use, payload and recovery from disturbances. The preprint's combined hardware-and-control approach is promising because it treats embodiment as part of the learning system. Its larger value will be proven when other teams can build the robot, run the same evaluations and discover where the design succeeds or needs revision.