RobKiNet: Robotic Kinematics Informed Neural Network for Optimal Robot Configuration Prediction

TL;DR

The paper tackles the bottleneck at the task–motion planning interface in TAMP for robots with redundant DOFs by introducing RobKiNet, a neural framework that embeds robotic kinematics as differentiable constraints. It presents two predictors, CMP for chassis-level motion and FMP for whole-body control, trained with differentiable IK/FK engines to produce kinematically feasible configurations directly from task goals. Empirical results show CMP achieves $ACC_{CMP}=96.67\%$ and FMP achieves $ACC_{FMP}=98\%$, with order-of-magnitude speedups in motion planning and exceptional data efficiency ($1/71$ for CMP and $1/15052$ for FMP) compared with traditional DL and DRL baselines. Real-device experiments on disassembly tasks confirm high reliability and near-1 mm end-effector accuracy, highlighting the practical impact of integrating differentiable kinematics into robotic TAMP for robust and efficient autonomous manipulation.

Abstract

Task and Motion Planning (TAMP) is essential for robots to interact with the world and accomplish complex tasks. The TAMP problem involves a critical gap: exploring the robot's configuration parameters (such as chassis position and robotic arm joint angles) within continuous space to ensure that task-level global constraints are met while also enhancing the efficiency of subsequent motion planning. Existing methods still have significant room for improvement in terms of efficiency. Recognizing that robot kinematics is a key factor in motion planning, we propose a framework called the Robotic Kinematics Informed Neural Network (RobKiNet) as a bridge between task and motion layers. RobKiNet integrates kinematic knowledge into neural networks to train models capable of efficient configuration prediction. We designed a Chassis Motion Predictor(CMP) and a Full Motion Predictor(FMP) using RobKiNet, which employed two entirely different sets of forward and inverse kinematics constraints to achieve loosely coupled control and whole-body control, respectively. Experiments demonstrate that CMP and FMP can predict configuration parameters with 96.67% and 98% accuracy, respectively. That means that the corresponding motion planning can achieve a speedup of 24.24x and 153x compared to random sampling. Furthermore, RobKiNet demonstrates remarkable data efficiency. CMP only requires 1/71 and FMP only requires 1/15052 of the training data for the same prediction accuracy compared to other deep learning methods. These results demonstrate the great potential of RoboKiNet in robot applications.

Figures (10)

• Figure 1: Overview of this paper's issues, ideas, chapter relationships, and main contributions.
• Figure 2: Example—Simplified TAMP framework for disassembling a screw on an EOL-EVB. (Left) Task Planning Level: The task planner, based on operational requirements (e.g., disassembling a particular bolt), outputs a sequence of action primitives (such as Approach, Disassembly, etc.) that guide the robot from its current state to the goal state. (Middle) Interface Between Task Planning and Motion Planning: This essential interface uses sampling methods to search for the robot’s configuration parameters and outputs them to the motion planning level. (Right) Motion Planning Level: The robot’s configuration parameters predicted by the interface—such as chassis position and joint angles of the robotic arm—are used to ensure the feasibility of execution. In this study, the predicted parameters from the interface meet both the global constraints of task planning (e.g., ensuring the robot is positioned near the target screw during the Approach primitive) and significantly enhance the efficiency of subsequent motion planning (ensuring feasible kinematic solutions).
• Figure 3: Comparison of three robot configuration (such as the chassis position) sampling methods. (a) Randomized sampling. (b) NN Regression method. (c) RobKiNet proposed in this work. The effectiveness of sampling in a continuous solution space will greatly affect the efficiency of motion planning. RobKiNet eliminates the need to iterate several times to validate the kinematics in a simulation environment.
• Figure 4: An overview of differentiable kinematics and its integration with neural networks. (a) Overview of the joint coordinate systems for the 6-DOF robotic arm kinematics. (b) The computational process for the analytical inverse kinematics solution. (c) The computational process for the analytical forward kinematics solution. (d) Non-differentiable operations and function modules within the program. (e) The motivation behind this study. (f) The basic concept of differentiable programming — ensuring the chain rule of the computational graph. This is achieved by transforming non-differentiable operators (red circles) into differentiable operators (green). There are other challenges to differential programming (2, 3), and solutions to all these problems will be detailed later. (g) The integration of $\mathbb{IK}$ with neural networks for chassis position prediction (CMP). (h) The integration of $\mathbb{FK}$ with neural networks for predicting both chassis position and joint angle configuration (FMP).
• Figure 5: Differential Method summary, corresponding to the treatment of non-differentiable operators in differential kinematics(red circle in Figure \ref{['Overview of Differentiable Kinematics']}).