[Paper Notes] Self-Contained and Automatic Calibration of a Multi-Fingered Hand Using Only Pairwise Contact Measurements
Published:
This post supports English / 中文 switching via the site language toggle in the top navigation.
TL;DR
This paper presents a calibration procedure for a multi-fingered robotic hand that needs no camera, marker, fixture, or external reference object. The hand brings two fingertips into contact, records the joint configuration at first contact, and fits its kinematic parameters so the modeled distance between those fingertip geometries is zero. Repeating this process across every finger pair turns the hand into its own calibration instrument.
The main intellectual contribution is an observability argument. Each contact supplies only one scalar distance constraint, so it cannot recover every absolute kinematic quantity. The authors show that a joint calibration over multiple finger chains still identifies every parameter direction that affects the relative fingertip positions needed for grasping and in-hand manipulation. A single finger pair remains insufficient; adding a third chain removes the relevant ambiguity.
On the four-fingered DLR-Hand II, the method reduces maximum pairwise distance error from 17.70 mm to 3.69 mm and mean error from 6.07 mm to 0.72 mm. The propagated mean error for the manipulation-oriented relative-position metric falls from 8.01 mm to 0.89 mm. About 150 contacts, collected automatically in 9 minutes, are sufficient for accurate calibration.
Paper Info
“Self-Contained and Automatic Calibration of a Multi-Fingered Hand Using Only Pairwise Contact Measurements” is by Johannes Tenhumberg, Leon Sievers, and Berthold Bäuml. It was presented at the 2023 IEEE-RAS 22nd International Conference on Humanoid Robots (Humanoids). The paper is available from the DLR Electronic Library, arXiv:2311.03957, and IEEE. The authors also provide a project page.
1. Calibration from a Zero Measurement
A dexterous hand needs an accurate forward-kinematics model. Small geometric errors move fingertips away from planned contacts, weaken grasp planning, and widen the sim-to-real gap for tactile in-hand manipulation. Conventional calibration often tracks fingertip markers with an external camera or electromagnetic system. A compact multi-fingered hand makes that setup awkward because markers occlude one another and several end effectors share a small workspace.
The paper exploits a measurement already available inside the mechanism: two bodies are at zero separation when contact begins. Let the forward kinematics be
[ f(q,\rho)=F, ]
where (q) contains joint angles and (\rho) contains the Denavit–Hartenberg parameters. For a fingertip pair (u=(E_k,E_l)), the relative transform is
[ {}^{E_k}T_{E_l}=f(q,\rho){E_k}^{-1}f(q,\rho){E_l}. ]
Given the fingertip geometries, their signed or unsigned distance is a function of this relative pose:
[ h_c^u(q,\Theta)=d^u!\left({}^{E_k}T_{E_l}\right). ]
At the measured contact configuration (q^{(n)}), the target is always
[ y^{(n)}=0. ]
The data therefore reverse the usual sensing pattern. An external tracker selects (q) and returns a Cartesian position (y). Contact calibration already knows (y=0); the search motion discovers the joint configuration (q) at which that constraint becomes true.
All parameters are fitted jointly with a regularized maximum-a-posteriori least-squares objective:
[ \Theta^*=\arg\min_{\Theta} \left[ \sum_{n=1}^{N}\frac{|y^{(n)}-h(q^{(n)},\Theta)|^2}{\sigma_m^2} +(\Theta-\Theta_p)^\top\Lambda_p^{-1}(\Theta-\Theta_p) \right]. ]
The Gaussian prior stabilizes parameter directions that the measurements cannot identify. For the DLR-Hand II, fingertip contact regions are capsules, so distance is inexpensive to compute. A hand with mesh fingertips could use a collision-distance algorithm such as GJK.
2. Calibrate What Manipulation Actually Uses
The paper separates the measurement available on hardware from the quantity the downstream task needs. Dexterous manipulation depends primarily on relative fingertip positions. Choosing fingertip (E_1) as a reference gives
[ h_t^k(q,\rho)=f(q,\rho){E_k,x}-f(q,\rho){E_1,x}, \qquad k=2,\ldots,N_E. ]
A common translation applied to every fingertip cannot be recovered from pairwise contact, yet it also leaves these difference vectors unchanged. Requiring every DH parameter to be individually observable would therefore be unnecessarily strict.
The authors linearize the contact and task measurement functions with respect to calibration parameters:
[ J_s=\left.\frac{\partial h(q_s,\Theta)}{\partial\Theta}\right|_{\Theta_0}. ]
The nullspace of (J^\top J) contains parameter directions that a measurement cannot sense. Contact calibration has enough information for the manipulation task when
[ \operatorname{kernel}(J_c^\top J_c) \subseteq \operatorname{kernel}(J_t^\top J_t). ]
This inclusion is the conceptual center of the paper: every ambiguity left by contact must also be irrelevant to relative fingertip placement.
For the full DLR-Hand II, contact and task Jacobians both have 56 eigenvalues above (10^{-6}) among 64 DH parameters. The eight null directions come from parallel joint axes and do not alter the fingertip task metric. Their kernels satisfy the required inclusion.
The result depends on calibrating the hand as one kinematic tree. For a single pair, 28 parameter directions should matter, but the task measurement observes 27 and scalar contact observes only 26. Two fingertips can change relative orientation while remaining on a constant-distance sphere. With three fingers, this invariance disappears; using all six pairs in the four-fingered hand gives the strongest result. The practical rule is simple: collect contacts across the whole hand, then solve one joint calibration problem.
3. Choosing Informative Contacts
Contact imposes a severe sampling constraint: the hand can measure only configurations in which two fingertips touch. Those configurations cover a thin subset of joint space, while calibration quality must remain high across the full Cartesian workspace.
The authors address this distribution shift with task D-optimal experimental design. Candidate contacts are scored by the parameter covariance induced by contact measurements and by how that uncertainty propagates through the desired task Jacobians. In compact form, the criterion minimizes a determinant involving
[ \operatorname{cov}(\Theta) \quad\text{and}\quad \frac{1}{\bar N_D}\sum_s J_t^{s\top}J_t^s. ]
This score asks which feasible contacts will best reduce relative-fingertip error over a Cartesian-uniform test distribution. The paper compares greedy selection and DETMAX with random sampling.
In simulation, 100 perturbed hand models are generated with uniform noise of (\pm5^\circ) on rotational DH parameters and (\pm5) mm on translational parameters. Their initial average deviation is about 21 mm. Both optimized selection methods converge faster than random sampling, reaching roughly 0.1 mm mean error with 300 contacts. The advantage persists at larger dataset sizes because the contact and task distributions remain different.
4. Making Contact Collection Automatic
Each sample needs a safe search trajectory whose endpoint crosses the unknown real contact surface. The procedure is:
- Sample 100,000 configurations in the shared reachable workspace of a finger pair.
- Choose one configuration for finger A and find configurations of finger B whose modeled tips collide.
- Keep one finger passive; generate a start pose for the other finger that is clearly separated.
- Move along a path from modeled separation to modeled penetration and stop at detected contact.
- Move the two unused fingers away from the pair’s shared workspace and repeat for all six pairs.
This search is necessary because the uncalibrated error reaches 17.7 mm, roughly the fingertip size and around 10% of the hand workspace. A pose that appears to touch in the nominal model may miss completely on hardware.
DLR-Hand II has output-side torque sensing on all 12 active joints. Before an approach, the passive finger’s torque offset (\tau_0) is recorded. Contact is declared when the torque change exceeds 0.1 Nm, and the corresponding joint angles become the measurement. Detecting the earliest low-force contact reduces deformation and penetration error.
5. Real-Hand Results
The platform has four fingers, each with three active and one passive joint. Calibrating four DH parameters per joint yields 64 parameters. The authors collect 300 contact samples, use an 80/20 train–test split with cross-validation, and find that 150 samples are already sufficient. Automatic acquisition of those 150 samples takes nine minutes.
| Model | Contact mean (mm) | Contact std. (mm) | Contact max (mm) | Task mean (mm) |
|---|---|---|---|---|
| Nominal kinematics | 6.07 | 3.90 | 17.70 | 8.01 |
| Calibrated joint offsets | 1.04 | 0.82 | 5.13 | 1.36 |
| Calibrated full DH model | 0.72 | 0.58 | 3.69 | 0.89 |
Joint-offset calibration captures most of the easy correction. Fitting the full DH model still reduces the maximum residual by another 1.44 mm and the mean by 0.32 mm. The remaining 3.69 mm worst case is close to the “few millimeters” regime needed for tactile manipulation.
One evaluation detail deserves care. Contact errors are measured on held-out contact configurations. The task-space means in the last column are obtained by propagating contact uncertainty through the contact Jacobian, parameter covariance, and task Jacobian. They are therefore model-based estimates of relative-position uncertainty, not independent measurements from an external tracking system.
6. Strengths and Limitations
The strongest aspect is the alignment between calibration theory and the downstream task. The nullspace inclusion explains why losing absolute pose information is acceptable, and it reveals why all finger pairs must be calibrated together. The paper also covers the full operational loop: candidate generation, experimental design, safe approach motions, contact detection, nonlinear fitting, and real-hardware validation.
The method does assume accurate fingertip geometry and repeatable low-force contact detection. Soft fingertips, compliance, backlash, thermal drift, or contact hysteresis can shift the apparent zero-distance event. The current model fits rigid DH geometry on one hand design; elastic drivetrain and fingertip parameters are left for future work. A broader study across hand morphologies and sensing technologies would clarify generality.
The strongest numerical evidence is still contact-domain cross-validation. The task error is inferred through local linear uncertainty propagation, so direct external ground-truth measurements of relative fingertip positions would provide a useful complementary validation. Very small nonzero eigenvalues may also be practically fragile even when the theoretical rank condition holds.
Takeaway
This work turns self-contact into a metrology system. Its deeper lesson is that calibration should target task-relevant observability: a robot does not need to identify every geometric degree of freedom if the remaining ambiguities cannot change the quantities used by control.
For multi-fingered hands, scalar touch events become surprisingly informative when three ingredients are combined: all finger chains are solved jointly, contact poses are selected for their downstream information value, and uncertain nominal geometry is handled by active search motions. The result is a compact, automatic nine-minute routine that brings a 17.7 mm worst-case model error down to 3.7 mm without external equipment.