This article presents a comprehensive analysis of an impedance control strategy for redundant robotic manipulators, focusing on achieving high-precision tracking and compliant behavior for the end effector. The primary challenge in controlling redundant systems lies in effectively managing the extra degrees of freedom (DOF) without interfering with the main task execution at the end effector. We address this by formulating a control law that decouples the dynamics into task space and null space components, allowing for independent regulation of the end effector trajectory and the internal configuration of the arm.
Introduction and Problem Formulation
Redundant robotic manipulators, possessing more joints than required to define the pose of their end effector, offer significant advantages in dexterity, obstacle avoidance, and optimization of secondary criteria like energy consumption or joint limit avoidance. The fundamental problem is to control the position and orientation (pose) of the end effector along a desired trajectory while utilizing the redundancy to achieve auxiliary objectives or compliant behavior. A critical requirement in many applications, from assembly to human-robot collaboration, is for the end effector to interact safely and responsively with its environment. This necessitates control strategies that go beyond rigid trajectory tracking to incorporate force or impedance regulation.
The core difficulty stems from the non-unique mapping between joint space and task space. For a robot with \(n\) joints and an end effector task requiring \(m\) dimensions (e.g., \(m=6\) for full position and orientation), redundancy exists when \(n > m\). The set of joint motions that produce no motion at the end effector is defined as the null space of the kinematic Jacobian. The control objective is twofold: first, to ensure accurate tracking of the desired end effector trajectory in the task (Cartesian) space; and second, to implement a desired dynamic behavior, such as impedance control, within the null space to manage internal motions and external interactions without perturbing the primary task.
Dynamic Modeling of a Redundant Manipulator
The dynamics of an \(n\)-joint robotic manipulator in joint space are given by the Lagrangian formulation:
$$
B(q)\ddot{q} + C(q, \dot{q})\dot{q} + g(q) = \tau – \tau_{ext}
$$
where \(q, \dot{q}, \ddot{q} \in \mathbb{R}^n\) are the joint position, velocity, and acceleration vectors, respectively. \(B(q) \in \mathbb{R}^{n \times n}\) is the symmetric, positive-definite inertia matrix. \(C(q, \dot{q}) \in \mathbb{R}^{n \times n}\) is the Coriolis and centrifugal matrix. \(g(q) \in \mathbb{R}^n\) is the gravitational torque vector. \(\tau \in \mathbb{R}^n\) is the control torque input, and \(\tau_{ext} \in \mathbb{R}^n\) represents external torques acting on the joints.
The differential kinematics relating joint velocities to end effector velocities (linear velocity \(v_p\) and angular velocity \(\omega\)) is:
$$
v = \begin{bmatrix} v_p \\ \omega \end{bmatrix} = J(q)\dot{q}
$$
where \(J(q) \in \mathbb{R}^{m \times n}\) is the geometric Jacobian matrix. For a redundant manipulator (\(n > m\)), the general solution for \(\dot{q}\) is:
$$
\dot{q} = J^{\#}(q)v + N(q)\dot{q}_N
$$
Here, \(J^{\#}(q)\) is a generalized inverse of \(J(q)\) (e.g., the Moore-Penrose pseudoinverse), \(N(q) = I – J^{\#}(q)J(q)\) is a projector onto the null space of \(J(q)\), and \(\dot{q}_N \in \mathbb{R}^n\) is an arbitrary vector representing joint motion within the null space. To parameterize this internal motion, we introduce \(r = n – m\) auxiliary variables \(\lambda \in \mathbb{R}^r\) such that \(N(q)\dot{q}_N = Z(q)\dot{\lambda}\), where \(Z(q)\) is a full-rank matrix whose columns span the null space of \(J(q)\). A dynamically consistent generalized inverse for \(Z(q)\) is \(Z^{\#}(q) = (Z(q)^T B(q) Z(q))^{-1} Z(q)^T B(q)\). This allows us to define an extended task vector \(\xi = [v^T, \lambda^T]^T\) and an extended Jacobian \(J_E(q)\):
$$
\xi = \begin{bmatrix} v \\ \lambda \end{bmatrix} = J_E(q) \dot{q}, \quad \text{where} \quad J_E(q) = \begin{bmatrix} J(q) \\ Z^{\#}(q) \end{bmatrix}
$$
Assuming \(J(q)\) is full rank, \(J_E(q)\) is square and invertible. Its inverse is given by \(J_E^{-1}(q) = [J^{\#}(q) \quad Z(q)]\). The joint accelerations can then be expressed as:
$$
\ddot{q} = J_E^{-1}(q) \dot{\xi} + \dot{J_E}^{-1}(q) \xi
$$
The control law for the joint torque is designed as:
$$
\tau = B(q)u_q + C(q, \dot{q})\dot{q} + g(q)
$$
where \(u_q\) is a new virtual control input with the dimension of joint acceleration. Substituting this into the dynamic model yields the closed-loop joint space dynamics:
$$
\ddot{q} = u_q – B(q)^{-1}\tau_{ext}
$$
By choosing \(u_q\) appropriately, we can project the dynamics onto the task space and null space. Let us define virtual control inputs for the task space \(u_v\) and the null space \(u_{\lambda}\) such that:
$$
u_q = J^{\#}(q)(u_v – \dot{J}(q)\dot{q}) + Z(q)(u_{\lambda} – \dot{Z}^{\#}(q)\dot{q})
$$
Substituting and premultiplying by \(J(q)\) and \(Z^{\#}(q)\) respectively, we obtain the decoupled closed-loop dynamics:
Task Space Dynamics:
$$
\dot{v} = u_v – J(q)B(q)^{-1}\tau_{ext}
$$
Null Space Dynamics:
$$
\dot{\lambda} = u_{\lambda} – Z^{\#}(q)B(q)^{-1}\tau_{ext}
$$
This decoupling is fundamental. It shows that external torques \(\tau_{ext}\) affect both the primary task (via \(J(q)B(q)^{-1}\tau_{ext}\)) and the null space motion (via \(Z^{\#}(q)B(q)^{-1}\tau_{ext}\)). The controller design now focuses on defining \(u_v\) and \(u_{\lambda}\) to achieve precise end effector tracking and desired null space impedance.
Impedance Controller Design
The control architecture is designed with two main components: a primary task-space controller for the end effector and an impedance controller acting in the null space.
Primary Task-Space Controller
The primary objective is for the end effector to track a desired pose trajectory \(p_d(t)\), \(R_d(t)\), with associated desired velocities \(v_d\) and accelerations \(a_d\). Let \(p_e\) and \(R_e\) be the current position and orientation. We define the position error as \(p_{\epsilon} = p_d – p_e\). For orientation, we use a non-minimal representation to avoid singularities. Let \(R_{\epsilon} = R_d^T R_e\) be the error rotation matrix, parameterized by a suitable set of parameters \(\alpha\). The orientation error vector can be defined as \(o_{\epsilon} = f_o(\alpha)\), with its derivative satisfying \(\dot{o}_{\epsilon} = L(\alpha)\omega_{\epsilon}\), where \(\omega_{\epsilon} = \omega_e – \omega_d\) is the angular velocity error and \(L(\alpha)\) is non-singular.
The composite task error vectors are:
$$
e_t = \begin{bmatrix} p_{\epsilon} \\ o_{\epsilon} \end{bmatrix}, \quad e_v = \begin{bmatrix} \dot{p}_{\epsilon} \\ \omega_{\epsilon} \end{bmatrix}
$$
To achieve tracking and compensate for the disturbance from \(\tau_{ext}\) in Eq. (5), the task-space control input \(u_v\) is designed as:
$$
u_v = a_d + D_v e_v + K_v e_t – J(q)B(q)^{-1}\gamma
$$
where \(D_v\) and \(K_v = \text{diag}(K_p, K_o)\) are positive definite diagonal gain matrices. The term \(\gamma\) is an estimate of the external torque disturbance, generated by the following update law:
$$
\dot{\gamma} = -K_I \gamma + \tau_{ext} + K_{\gamma}^{-1} B(q)^{-1} J(q)^T e_v
$$
Here, \(K_I\) and \(K_{\gamma}\) are positive definite diagonal matrices. In practice, \(\tau_{ext}\) is often not directly measured. However, \(\gamma\) can be computed indirectly by integrating the above law, which effectively provides a dynamic estimate of the external torque. Substituting this control law into the task-space dynamics (5) yields the closed-loop error dynamics:
$$
\dot{e}_v + D_v e_v + K_v e_t = J(q)B(q)^{-1} e_{\gamma}
$$
where \(e_{\gamma} = \gamma + \tau_{ext}\). With proper tuning of \(K_I\), \(e_{\gamma}\) converges to zero, leading to stable end effector tracking.
Null Space Impedance Control
The null space motion, parameterized by \(\lambda\), is used to implement a compliant behavior without affecting the end effector trajectory. The goal is to make the internal dynamics of the manipulator behave like a desired mass-spring-damper system (impedance) in response to internal forces or to maintain a preferred joint configuration \(q_d\).
We define the null space velocity error as \(e_{\lambda} = \dot{\lambda}_d – \dot{\lambda}\) (where \(\dot{\lambda}_d\) is often zero) and the joint configuration error as \(e_q = q_d – q\). The null space control input \(u_{\lambda}\) is designed to achieve the following impedance relationship:
$$
\Lambda_{\lambda}(q) \dot{e}_{\lambda} + \mu_{\lambda}(q, \dot{q}) e_{\lambda} + D_{\lambda} e_{\lambda} + Z(q)^T (K_q e_q + D_q \dot{e}_q) = Z(q)^T \tau_{ext}
$$
where \(\Lambda_{\lambda}(q) = Z(q)^T B(q) Z(q)\) is the null space inertia matrix, \(\mu_{\lambda}(q, \dot{q}) = Z(q)^T C(q, \dot{q}) Z(q) – \Lambda_{\lambda}(q) \dot{Z}^{\#}(q) Z(q)\) represents null space Coriolis/centrifugal terms, and \(D_{\lambda}\), \(K_q\), \(D_q\) are positive definite damping and stiffness matrices to be tuned. The right-hand side, \(Z(q)^T \tau_{ext}\), is the projection of external joint torques onto the null space.
To realize this impedance, the null space virtual control input is chosen as:
$$
u_{\lambda} = \ddot{\lambda}_d + \Lambda_{\lambda}(q)^{-1} \left( \mu_{\lambda}(q, \dot{q}) e_{\lambda} + D_{\lambda} e_{\lambda} + Z(q)^T (K_q e_q + D_q \dot{e}_q) \right)
$$
The figure illustrates the concept of null space motion. The primary motion of the end effector from A to B is achieved by specific joint movements. However, redundancy allows for an infinite number of joint trajectories to achieve the same end effector path. The null space impedance controller shapes this internal motion, for instance, to keep joints near the center of their range (Trajectory 2) or to minimize kinetic energy, without altering the path of the end effector itself.
Substituting this \(u_{\lambda}\) into the null space dynamics (6) directly yields the desired impedance equation (13). This equation shows that if an external torque \(\tau_{ext}\) has a component in the null space, it will excite the impedance dynamics, causing a compliant displacement in \(\lambda\) and consequently in the joint configuration \(q\), while the primary task of the end effector remains undisturbed due to the decoupling.
Stability Analysis
The stability of the overall closed-loop system can be proven using Lyapunov’s direct method. Consider the composite system state defined by the errors: \(x = (e_q, e_t, e_v, e_{\gamma}, e_{\lambda})\).
Theorem: For the redundant manipulator dynamics (1) under the control laws (9), (10), (12), and (14), with positive definite gain matrices \(K_v, D_v, K_I, K_{\gamma}, K_q, D_q, D_{\lambda}\), and assuming bounded desired trajectories and external torques (\(\dot{\tau}_{ext}=0\)), the following holds:
- If \(\tau_{ext}=0\) and \(q_d\) is chosen such that the end effector can achieve the desired pose \((p_d, R_d)\), then the origin \(x=0\) is asymptotically stable.
- If \(\tau_{ext} \neq 0\), the state \(x\) converges to an equilibrium where the end effector pose error \((e_t, e_v)\) is zero, the disturbance estimate error \(e_{\gamma}\) is zero, and the joint configuration \(q\) converges to a value \(q^*\) that locally minimizes the quadratic function \(\| K_q e_q + D_q \dot{e}_q – \tau_{ext} \|^2\) subject to the end effector being at \((p_d, R_d)\).
Proof Sketch: A candidate Lyapunov function is constructed in parts. For the task-space dynamics, consider:
$$
V_1(e_v, e_t, e_{\gamma}) = \frac{1}{2} e_v^T e_v + \frac{1}{2} e_t^T K_v e_t + \frac{1}{2} e_{\gamma}^T K_{\gamma} e_{\gamma} + V_o(\alpha)
$$
where \(V_o(\alpha)\) is a positive definite function of the orientation parameters such that \(\dot{V}_o = -\omega_{\epsilon}^T (K_o L(\alpha)) o_{\epsilon}\). Using the closed-loop error dynamics (11), the derivative \(\dot{V}_1\) can be shown to be negative semi-definite: \(\dot{V}_1 = -e_v^T D_v e_v – e_{\gamma}^T K_{\gamma} K_I e_{\gamma} \leq 0\).
For the null space dynamics, consider the Lyapunov function within the invariant set where \(e_v=0, e_t=0, e_{\gamma}=0\):
$$
V_2(e_{\lambda}, e_q) = \frac{1}{2} e_{\lambda}^T \Lambda_{\lambda}(q) e_{\lambda} + \frac{1}{2} e_q^T K_q e_q
$$
Its derivative, using the impedance equation (13) and the property \(\dot{\Lambda}_{\lambda}(q) – 2\mu_{\lambda}(q, \dot{q})\) being skew-symmetric, is:
$$
\dot{V}_2 = -e_{\lambda}^T (D_{\lambda} + Z(q)^T D_q Z(q)) e_{\lambda} – e_{\lambda}^T Z(q)^T \tau_{ext}
$$
With \(\tau_{ext}=0\), \(\dot{V}_2 \leq -e_{\lambda}^T D_{\lambda} e_{\lambda} \leq 0\). Applying LaSalle’s invariance principle proves asymptotic convergence of both task-space and null-space errors, establishing the first result. The second result follows from analyzing the equilibrium conditions of the coupled system, showing that the null space impedance law minimizes the impact of external torques on the internal configuration while maintaining perfect end effector tracking.
Simulation Analysis and Performance Comparison
To validate the proposed impedance control scheme, simulations were conducted for a 7-DOF redundant manipulator and compared against a conventional decentralized joint-space PID controller. The primary task was for the end effector to track a complex spatial trajectory. The performance metrics were tracking error for the end effector and the smoothness of joint motion.
Control Parameters:
| Gain Matrix | Value | Description |
|---|---|---|
| \(K_p\) | \(500 I_3\) | Task position stiffness |
| \(K_o\) | \(500 I_3\) | Task orientation stiffness |
| \(D_v\) | \(\text{diag}(30I_3, 300I_3)\) | Task damping |
| \(K_I, K_{\gamma}\) | \(150 I_7\) | Disturbance estimator gains |
| \(K_q\) | \(40 I_7\) | Null space joint stiffness |
| \(D_q\) | \(10 I_7\) | Null space joint damping |
| \(D_{\lambda}\) | \(5 I_5\) | Null space impedance damping |
Comparative Results: The table below summarizes the key performance indicators comparing the proposed Impedance Control and a standard PID control for the end effector trajectory tracking task.
| Performance Metric | PID Control | Impedance Control | Improvement |
|---|---|---|---|
| RMS Position Error (m) | \(8.7 \times 10^{-3}\) | \(2.1 \times 10^{-3}\) | ~76% reduction |
| Peak Position Error (m) | \(2.4 \times 10^{-2}\) | \(5.8 \times 10^{-3}\) | ~76% reduction |
| RMS Orientation Error (rad) | \(5.1 \times 10^{-2}\) | \(1.3 \times 10^{-2}\) | ~75% reduction |
| Joint Torque Smoothness (\(\|\Delta\tau\|\)) | High | Low | Reduced chattering |
| Null Space Utilization | None / Unstructured | Active Compliance / Optimization | Enabled auxiliary tasks |
| Settling Time after Disturbance (s) | > 3.0 | < 1.5 | > 50% faster |
The simulation results clearly demonstrate the superiority of the null space impedance control. The PID controller, while attempting to track the desired end effector trajectory directly in joint space, shows significant error, especially during sharp directional changes. This is because it does not explicitly account for the dynamic coupling between joints or the task-space geometry. In contrast, the proposed controller’s task-space component ensures minimal tracking error for the end effector. Furthermore, the null space impedance behavior actively dampens internal oscillations and provides a compliant response. When an external force was applied to a link (simulating a collision), the PID controller caused the entire arm to oscillate, disturbing the end effector path. The impedance controller, however, absorbed the impact primarily in the null space, allowing the end effector to maintain its trajectory with only a minor, quickly corrected deviation.
The mathematical expression for the end effector tracking error norm over time \(t\) under the two controllers highlights this difference. Let \(e_{pos}(t) = \|p_d(t) – p_e(t)\|\). Empirically from the simulation data, the error envelope can be approximated as:
PID: $$ e_{pos}^{PID}(t) \approx A_{pid} e^{-\zeta_{pid} \omega_{n_{pid}} t} \sin(\omega_{d_{pid}} t + \phi) + b_{pid} $$
Impedance: $$ e_{pos}^{Imp}(t) \approx A_{imp} e^{-\zeta_{imp} \omega_{n_{imp}} t} $$
where \(A_{imp} << A_{pid}\), \(\zeta_{imp} \omega_{n_{imp}} > \zeta_{pid} \omega_{n_{pid}}\), and \(b_{pid}\) is a steady-state bias for PID. The impedance controller shows a faster, overdamped convergence due to its model-based compensation and disturbance rejection.
Discussion and Conclusion
The implementation of impedance control within the null space of a redundant manipulator provides a powerful framework for sophisticated robotic interaction tasks. The key achievement is the decoupling of the end effector‘s primary objective from the internal configuration management of the arm. This allows the end effector to perform precise, dynamic tracking while the arm simultaneously exhibits compliant, spring-damper-like behavior in its redundant directions. This compliance is crucial for operational safety, energy absorption during unexpected contact, and for executing secondary tasks like optimizing manipulability or avoiding joint limits.
The proposed control law offers several advantages for end effector operation, summarized below:
| Advantage | Impact on End Effector Performance |
|---|---|
| Dynamic Decoupling | External disturbances are projected and handled in the null space, minimizing their effect on the primary end effector trajectory. |
| Active Compliance | The end effector can maintain contact with a surface with a desired force profile by allowing compliant null space motion. |
| Enhanced Stability | The Lyapunov-based design guarantees stable convergence of both task-space and null-space errors. |
| Disturbance Rejection | The integrated estimator (\(\gamma\)) actively compensates for unmodeled dynamics and external loads, improving end effector accuracy. |
| Flexible Prioritization | The framework naturally allows for prioritization; the end effector task is always primary, with null space tasks being secondary. |
In conclusion, this article has detailed a systematic approach to impedance control for redundant manipulators by leveraging null space projections. The method ensures high-precision control of the end effector while embedding a tunable compliant behavior for the arm’s internal dynamics. The stability is rigorously proven, and simulation studies confirm significant improvements in tracking accuracy, disturbance rejection, and motion smoothness compared to traditional PID strategies. This makes the approach highly suitable for advanced robotic applications where the end effector must interact dynamically and safely with complex, uncertain environments.