← Paper Digest

Learning Accurate End-Effector Trajectory Tracking with Inertial Feedback for a Quadruped Manipulator

Dohyeon Kim, Seungyong Han, Yeongtae Jung

November 2026

TL;DR An IMU on the gripper of a Unitree Go2 with a PiPER arm lets a residual RL policy correct arm tracking errors that joint encoders cannot see, such as base motion and backlash. On hardware, motion-capture tracking error averages 16.42 mm, against 20.13 mm for the nominal controller and 18.26 mm for a residual policy without the IMU.

What they do

On a legged manipulator, errors from foot contact, backlash, compliance and calibration add up along the leg-arm chain where joint encoders cannot see them. The authors mount an IMU on the end effector of a Unitree Go2 carrying a 6-DoF PiPER arm. A residual RL policy uses its readings to correct the arm commands while the body and feet stay in place.

How it works

  • Residual arm policy. Kinematics-based leg and arm controllers stay in charge; a 50 Hz PPO policy adds joint-position corrections to the six arm joints only. Besides proprioception it observes the end-effector IMU’s gravity direction, angular velocity and linear acceleration.
  • Training. In Isaac Lab on 1,000 random multi-segment trajectories in a local workspace, with randomized noise, backlash and leg damping.
  • Two-stage sim-to-real alignment. CMA-ES fits arm joint stiffness and damping to real chirp responses, then a neural residual model makes the simulated IMU signals match real ones more closely. Without it the IMU policy oscillated badly on hardware.
  • Evaluation. Motion capture measures the real end effector but is never fed to the controller.

Results

  • Simulation, eight command conditions: the IMU policy has the lowest error in all eight. Its mean RMSE is 30.5% below the nominal controller and 5.5% below a residual policy without the IMU.
  • Hardware, in-distribution commands: motion-capture RMSE falls 17.3% and 28.0% versus the nominal controller, and 12.6% and 15.9% versus the policy without the IMU.
  • Averaged over all eight hardware cases, RMSE is 20.13 mm for the nominal controller, 18.26 mm without the IMU and 16.42 mm with it.
  • Forward kinematics and motion capture disagree by 12.61 ± 2.15 mm RMSE across 24 runs.

Limitations

The policy only tracks positions in a small workspace with orientation held fixed, and never moves the body or feet. Under the fastest out-of-distribution commands it is no better than the policy without the IMU (0.6% worse in one case), which the authors attribute to actuator bandwidth and the policy rate. Whole-body motion and other robots are left to future work.

Figures

From Figure 1: Nominal leg and arm controllers produce baseline commands, and a residual RL policy for the arm adds c...
From Figure 1: Nominal leg and arm controllers produce baseline commands, and a residual RL policy for the arm adds corrections using proprioception and end-effector inertial feedback. Sim-to-real alignment combines simulation parameter identification with neural residual modeling of IMU discrepancies.
From Figure 3: The hardware: a Unitree Go2 quadruped with a PiPER arm, an IMU on the end effector, and motion-capture...
From Figure 3: The hardware: a Unitree Go2 quadruped with a PiPER arm, an IMU on the end effector, and motion-capture markers on the end effector and the base.
From Figure 2: Sim-to-real alignment. (a) Simulated and real arm joint trajectories under chirp excitation after para...
From Figure 2: Sim-to-real alignment. (a) Simulated and real arm joint trajectories under chirp excitation after parameter identification. (b) Frequency spectra of real and simulated end-effector IMU signals, with and without the neural residual model.
Abstract

Accurate end-effector trajectory tracking on quadruped manipulators is challenging because uncertainties in actuators, foot-ground contact, and geometric errors are propagated through the leg-arm kinematic chain. Detecting such errors requires sensing beyond joint encoders that can capture the effects of these uncertainties on the robot's motion, since joint encoders only provide measurements at the actuated joints and cannot fully observe mechanical errors such as backlash, dimensional inaccuracies, or structural deformation. This work addresses these challenges using a reinforcement learning framework with end-effector (EE) inertial feedback, targeting local high-accuracy manipulation while maintaining a nominal body pose and support configuration without intentional body motion or foot repositioning. The learned policy uses EE inertial measurements to compensate for tracking errors induced by uncertainties throughout the leg-arm system that cannot be fully inferred from proprioception alone. Simulation and real-world experiments on continuous EE trajectory tracking show overall improvements relative to nominal and proprioception-only baselines. In real-world experiments, the proposed policy reduces the MoCap-based RMSE by 17.3% and 28.0% relative to the nominal baseline under the two in-distribution (ID) command conditions, respectively, and by an additional 12.6% and 15.9% relative to the proprioception-only baseline. Furthermore, the trajectory-wise RMSE between FK-reconstructed and motion-capture-measured end-effector positions averaged 12.61+-2.15mm (mean+-standard deviation) across 24 controller-trajectory combinations, revealing a substantial discrepancy between nominal kinematic reconstruction and measured end-effector motion.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.