ONGOING LAB RESEARCH · QUADRUPED ROBOTICS
面向陡峭地形的 15 自由度主动脊柱仿生四足机器人
TerraGoat is a goat-inspired research platform for steep slopes and sparse footholds. An active spine changes body configuration and center-of-mass distribution, while interchangeable feet, modular hardware, and a ROS 2 foundation support research into coordinated spine–limb climbing.
Internal laboratory project
Source code and internal documentation cannot be shared. This page presents only approved system design, research goals, and project-stage information.
Rigid-trunk quadrupeds must rely almost entirely on their limbs for center-of-mass transfer and foot repositioning on steep terrain. TerraGoat makes the body itself a controllable kinematic chain, asking whether active-spine motion can expand reachable workspace, preserve contact, and improve resistance to tipping.
15 DOF
Limbs and active spine
ROS 2
Modular control stack
3D Print
Rapid structural iteration
RL
Spine–limb roadmap
The platform combines modular joints with additive-manufactured structures for rapid iteration around steep-contact tasks. The design treats spine deformation, foot contact, and cable routing as quantities that must later enter control and identification—not as isolated mechanical details.
Controllable joints between the front and rear body sections adjust trunk curvature, reachable foot workspace, and center-of-mass location.
3D printing accelerates geometry tests, while infill and engineering materials are tuned at load-critical parts to balance stiffness and mass.
Communication and power routes are integrated into the mechanism to reduce entanglement through spine and leg motion while preserving service access.
Split-toe geometry and contact pads test stepping/catching and spread-grip modes while supporting friction calibration.
Offline policy learning is separated from physical execution. Simulation handles physics alignment, domain randomization, and training; the host runs ROS 2 motion/spine nodes and ros2_control, maps target joint angles to 15 servos, and reads IMU and joint feedback.
OFFLINE
Isaac Lab / MuJoCo · PyTorch · PPO
HOST PC
Raspberry Pi / Jetson · Ubuntu · ROS 2
CONTROL
Hardware interface · Target joint angles
ROBOT
15× servos · OpenCR IMU · Joint states
PPO / PyTorch
Policy training
ROS 2 / OpenCR
Runtime interface
IMU / Joint State
State feedback
The project is currently in the physical-platform buildout and control-foundation phase. The 15-DOF active-spine prototype, wiring and hardware layout, ROS 2/URDF model, and open-loop motion tests are complete; upcoming work focuses on simulation–physics alignment and PPO-based spine–limb climbing control.
The active spine, limbs, modular joints, CAD assembly, fabrication, and physical integration are in place.
Servo, communication, power, and onboard-controller layouts have been assembled with cable-entanglement iterations.
The modular ROS 2 node structure, hardware interface, URDF/Xacro model, and initial kinematic/dynamic derivations have been established.
Open-loop gait and joint-target tracking tests provide data for action mapping, response-delay measurement, and state-noise analysis.
PPO-based spine–limb climbing control is the next stage and remains in planning and learning-environment preparation.
Physical experiments primarily establish credible parameter ranges for the future learning environment; they do not yet demonstrate autonomous climbing.
Measure angle/velocity command tracking across 15 joints, actuator response delay, and usable output boundaries.
Characterize IMU attitude/rate, joint feedback, and the influence of printed-spine deformation on center-of-mass estimation.
Compare measured joint angles with the URDF/Xacro model and inspect coordinated spine–limb workspace.
Measure foot-slip thresholds on the slope and estimate mass-distribution error, cable tension, and spine damping ranges.
The planned PPO policy coordinates active spine and limb motion rather than replacing all low-level control. It outputs target joint angles for execution by lower-level PD and hardware controllers.
Rebuild the robot and steep terrain in Isaac Lab or MuJoCo, injecting measured IMU noise, motor delay, mass error, and contact parameters into domain randomization.
Observe IMU, 15-DOF joint positions/velocities, foot contacts, and spine pose; define actions as target angles for low-level joint controllers.
Balance climbing progress, center-of-mass stability, contact retention, and energy/torque penalties.
Run ablations and robustness tests in parallel simulation before staged hardware deployment and external pose/trajectory validation.