Jumbot-B: A 3D-Printed Bipedal Robot

By Zhenkai Gao & Xianmai (Chris) Liang

Course: ME 134 — Robotics Studio, Spring 2026, Tufts University

A 3D-printed bipedal robot built end-to-end: CAD, fabrication, electronics, and gait tuning. Jumbot-B walks at a best speed of 4 cm/sec using manually-tuned periodic gaits derived from servo encoder data.

Walking video → · Modular leg-swap demo → · All build & test footage (Drive folder) →

Overview

Jumbot-B is a small bipedal robot driven by eight servos and a Raspberry Pi, with chassis and feet 3D-printed in PLA. Each leg has hip pitch + hip yaw + hip roll plus a knee/leg joint — enough articulation to shift weight and twist the body during single-support phases of a gait.

We tuned gaits directly on the hardware using servo encoder feedback rather than running a simulation track.

Robot Specifications

Robot Design & CAD

Key components: Raspberry Pi 4 compute on top, bus controller + DC converter on the platform, hip pitch / yaw / roll servos per leg, knee pitch servo, 3D-printed feet, and an IMU + battery on the underside.

Key design choices:

• DOFs (per leg): hip pitch + hip yaw + hip roll + knee/leg.
• Materials: PLA for the chassis and feet — rigid, predictable contact dynamics.
• Sensing: each servo reports position, velocity, temperature, and bus voltage; an onboard IMU provides body-frame orientation; encoder data drives the gait tuning loop.
• Compute: a Raspberry Pi 4 runs the gait controller and commands the servos over a half-duplex bus.

Main dimensions: 374 mm tall, 230 mm wide, 140 mm front-to-back, 122 mm foot stance.

Multi-pose CAD Views

Isometric, side, and front views.

Top, iso-rear, rear, and side views — chassis shown semi-transparent so the internal electronics layout is visible.

Mechanical Assembly

The robot was built outward from a single-leg prototype: Onshape CAD → 3D-printed parts (PLA, Prusa MK4S for v1.0 prototyping, Prusa Core One for v2.0 final parts) → support-structure removal → assembly → single-leg motion test → tandem-leg test → full bipedal robot.

Single-leg motion test video →

From CAD to real world — the right-leg assembly in Onshape (left) and the assembled 3D-printed leg with servos installed (right).

3D printing process — parts oriented in the slicer with overhangs facing up, v1.0 printing on a Prusa MK4S, v2.0 final parts on a Prusa Core One, and the printed parts with support structures cleaned off.

Exploded view of the full assembly — chassis plate, four servos per leg, linkage parts, hardware, and the foot module.

Three perspective renders of the assembled robot.

Assembled Robot Photos

Painted/sanded final assembly on the bench — chassis bolted up, all 8 servos wired into the bus, cable bundle routed along the spine.

Behind the scenes: standing on the lab floor between gait-tuning runs.

Failure & Iteration Log

Issues found during the single-leg build phase and how each was resolved:

Failures captured at the hardware: (1) the cracked cyan hip mount split along the print layers under shear; (2) the original foot wouldn’t seat onto the bottom leg link — mouse for scale; (3) the snapped servo disk in-situ after the linkage drove it past its mechanical limit.

Bipedal-stage iterations (post-assembly):

• TPU foot experiment. We replaced PLA feet with TPU (85A) hoping for better ground compliance. The flexibility expanded the gait search problem — the robot lost static balance and was harder to control — and we reverted to PLA.
• Power delivery upgrade. The original thin wiring between the DC converter and battery produced enough voltage drop to make servo behavior inconsistent. Replacing it with thicker-gauge cables stabilized the power rail and gait reliability immediately improved.
• Cable routing. Wiring tidied and bundled along the chassis so legs can swing without snagging.

Hardware Implementation

The robot is controlled by a Raspberry Pi running a periodic gait function. Servo commands go out directly over GPIO; encoder positions are read back to record waypoints during tuning. Power comes from an onboard battery via a DC converter.

Walking video: Jumbot-B taking baby steps

Frame sequence captured from the walking video — body shifts side-to-side as the swing leg clears the floor between consecutive support phases.

Hip Kinematics

Three hip degrees of freedom — yaw, roll, and pitch — give Jumbot-B body-twisting capability for shifting weight during single-support phases of a gait.

Knee Linkage

Four-bar knee linkage driven by the leg servo. Servo disk operating range: 133° (extended) to 290° (flexed). The mechanical limits were validated empirically by sweeping the servo through its full range and noting where the linkage starts to bind.

Servo Safety Limits

Per-joint position bounds, validated empirically by sweeping each servo through its full physical range and noting where the linkage starts to bind:

The gait controller clamps every servo command against these bounds so a misbehaving function can’t drive a servo into the chassis or fight the linkage.

Performance & Learning Curve

Speed measurements across iterations:

That’s roughly a 3.1× speedup from first walking gait to final, driven by hardware reliability fixes (thicker power wiring, cable management) and additional gait waypoint tuning.

Speed measurement at the Baby Steps milestone — robot walking next to a tape measure: 1 in/sec at peak, 2 in/cycle, 2 sec/cycle.

Gait Development

Our approach to gait generation:

1. Manually pose the robot through one walking cycle, capturing joint angles at each transitional state via the servo encoders.
2. Use those waypoints to construct a periodic function for the joint trajectories.
3. Iterate by hand, adjusting timing and amplitude until the robot walks stably.

The hip yaw + roll axes give Jumbot-B more body-twisting capability — useful for shifting the center of mass during single-support phases — but they also widen the gait search space, which is the main reason gait tuning has been the slow part of the project. Once roughly half a walking cycle is dialed in, the other half tends to mirror it.

Motor Data

We log per-servo position, velocity, temperature, and bus voltage during a walking cycle. The figure below covers an 8-second window across all 8 servos (Hip Pitch 1/2, Hip Yaw 1/2, Hip Roll 1/2, Leg 1/2):

• Position (deg): sinusoidal trajectories with leg/hip-pitch joints traversing the largest range; the periodic structure is exactly what the gait function commands.
• Velocity (deg/s): derivative of position; peaks near ±90 deg/s for the swing-leg joints.
• Temperature (°C): servos warm steadily from ~30 °C to ~40 °C over the run — comfortably below the failure window.
• Bus voltage (mV): ~8400–8800 mV with the upgraded wiring; pre-upgrade dips that caused jitter are no longer present.

Servo data during the Baby Steps gait — 8 s of position, velocity, temperature, and bus voltage logged from all 8 servos.

Energy Efficiency

Quick efficiency numbers derived from the spec sheet and best-speed measurement:

Cost of Transport is the standard dimensionless efficiency metric for legged locomotion. Humans walking sit around CoT ≈ 0.2; research quadrupeds like ANYmal land in the 1–3 range. Jumbot-B’s ~2.0 is in the right ballpark for a small experimental biped, though there’s clear room to optimize — most of the 5.4 W is going to support-leg holding torque, not forward motion.

Robot Reliability Routines

• Boot health check. Queries each servo for position, temperature, and bus voltage; aborts startup if any servo fails to respond or reports out-of-range temperature.
• Homing routine. Drives all 8 servos to a known stand-up pose via safe_move, which clamps every commanded angle to the per-servo bounds before issuing the move. Used at startup and as a recovery pose between gait runs.
• Servo-disconnect catch. The main control loop wraps servo commands in a try/except for ServoTimeoutError. If a motor disconnects or times out mid-gait, the loop logs a critical error and exits cleanly rather than charging on with a dead actuator.
• Shutdown routine. Detorques servos, returns the robot to the home pose, then powers down — prevents the robot from collapsing under its own weight when killed mid-run.

The homing() function: each leg’s four servos move to their home angle through a clamped safe_move wrapper, then a 2.5 s settle pause.

except ServoTimeoutError:
    print("\nCRITICAL ERROR: A motor disconnected or timed out!")
    print("Initiating emergency stop...")
    break

Lessons Learned

• Material choice has hidden second-order effects. TPU feet looked good on paper for compliance, but the flexibility blew up the gait problem and broke static balance.
• Power delivery is invisible until it isn’t. Thin wires produced enough voltage drop to make servo behavior inconsistent — symptoms looked like control bugs but were really hardware. Thick cables fixed it.
• Sim-to-real has a fixed setup cost. When that cost is high relative to the iteration budget — as it was once Onshape URDF export proved unreliable — tuning directly on hardware can dominate.
• Encoder-driven manual gaits work. The waypoint-based periodic function is crude but converges quickly once you find the first half of the cycle.
• Build quality before control quality. Three early single-leg failures (hip-mount cracking under shear, tight foot tolerances, servo-disk breakage from motion conflict) all came from CAD/print/assembly mismatches. Fixing each at the source — wall thickness, tolerance, motion-range checks — was much cheaper than working around them in software.

Future Work

• Refine the walking gait beyond the current 4 cm/sec.
• Land the Onshape → URDF pipeline so we can prototype gaits in MuJoCo before committing to hardware.
• Add new locomotion modes: turning, uneven terrain, jumping.
• Sweep a full power-vs-velocity efficiency curve to find the CoT-minimizing gait.

Bill of Materials

3D-printed structural parts (PLA except where noted) plus off-the-shelf hardware and electronics:

Code Repository

github.com/ChrisLiang33/jumbotB

Citations

[1] Sun, B., Mohamed Haris, S., & Ramli, R. (2026). A Systematic Review of Deep Reinforcement Learning for Legged Robot Locomotion. Instruments, 10(1), 8. https://www.mdpi.com/2410-390X/10/1/8

Acknowledgements

Built for ME 134 — Robotics Studio, Spring 2026 at Tufts University. Thanks to the course staff for the platform and feedback throughout the term.

Project by Zhenkai Gao and Xianmai (Chris) Liang.