Kinematics, Contact Mechanics, and Grasp Quality¶
The exporter preserves the link geometry used to generate contacts. The planner then maps each contact set to the wrenches it can exert on the object. This chapter defines both mappings and the lower-tail score used to compare planned approaches.
Joint Coordinates and Link Geometry¶
A hinge constrains the relative motion of two bodies to rotation about one axis. A serial finger chain composes these joint motions from the palm to the fingertip. Under the standard Denavit–Hartenberg convention, the transform from link frame \(i\) to its predecessor is
Here \(q_i^0\) is the MuJoCo reference angle, \(\theta_i^0\) is the DH angle at reference, \(d_i\) is the offset along the hinge axis, \(a_i\) is the common-normal length, and \(\gamma_i\) is the twist. The exporter constructs the DH frames from the compiled hinge axes at reference and transforms the participating collision meshes into their corresponding link frames. Geometry rigidly attached to a link shares its transform. GraspIt! XML stores angular constants in degrees and mesh coordinates in millimeters.
Some fingers contain an independent hinge and a driven hinge linked by an affine MuJoCo equality. If \(u_j=q_j\) is the driver coordinate, with reference angles \(q_i^0\) and \(q_j^0\), the equality and its reconstruction in absolute joint coordinates are
For \(q_j^0=0.35\) rad, \(q_i^0=-0.20\) rad, \(c_0=0.16\) rad, and \(c_1=0.70\), the manifest offset is \(b_i=-0.285\) rad. The manifest expresses the absolute-coordinate map in radians. The source convention is defined in the MuJoCo XML reference.
Frictional Point Contacts¶
Let \(p_i\) be the position of contact \(i\), \(n_i\) its inward unit normal, and \(c\) the object center of mass, all in the object frame \(\{O\}\). In the frictional point-contact model, the hand exerts a force \(f_i\) with normal component \(f_{n,i}=n_i^\top f_i\) and tangential component \(f_{t,i}=f_i-f_{n,i}n_i\). The unilateral Coulomb conditions are
A contact exerts a moment about \(c\) through the force lever arm. The force basis at a point has three components. The moment about \(c\) is therefore determined by \(f_i\) and \(p_i\). See Murray, Li, and Sastry, Chapter 5 and Rimon and Burdick, Chapter 4 for the grasp map and its contact models.
Grasp Wrench Space and Force Closure¶
The scorer approximates each cone by a polygonal cone with eight equally spaced tangent directions \(s_{ij}\). Each edge supports one unit of normal force. With moment scale \(\rho>0\), the edge forces, primitive wrenches, and normalized grasp wrench space are
A convex combination assigns nonnegative coefficients to the edge wrenches and constrains their sum to one. Since \(n_i^\top f_{ij}=1\), this constraint imposes a unit total normal-force budget. It is a contact-level normalization of wrench capacity. Force closure requires \(0\) in the interior of \(W\). If this condition holds, a sufficiently small wrench in every direction lies within the achievable set. See Roa and Suárez for the effect of force budgets and the moment scale on wrench-space metrics.
For a full-dimensional hull \(W=\{w:a_h^\top w\le b_h\ \forall h\}\) with \(0\) in its interior, the Ferrari–Canny metric is
This radius is the magnitude of the largest origin-centered Euclidean ball contained in \(W\). The Python planner uses \(\rho=\max(0.01\,\mathrm{mm},\max_i\lVert p_i-c\rVert_2)\) for \(\varepsilon\). Comparisons across grasp records must account for the moment normalization and friction coefficient. The recorded native GraspIt! epsilon is a separate quantity returned by the simulator.
Task-Wrench Capacity¶
For each loading condition \(k\), the planner receives a point \(p_k\) and the direction \(f_k\) of the contact force required to balance the external load. It normalizes \(\lVert f_k\rVert_2=1\) and forms the wrench \(\ell_k=(f_k,(p_k-c)\times f_k/\rho)\). The task score is the largest common load multiplier supported along all specified rays:
The task scorer requires a full-dimensional hull with the origin strictly inside. It returns zero when this force-closure condition fails. The planner uses \(\rho=100\,\mathrm{mm}\) for both the hull and the task wrenches. A common invertible rescaling of their moment coordinates preserves each ray intersection multiplier. For fixed contacts, increasing \(\mu\) expands each polygonal cone cross-section and hence \(W\). The supported multiplier is therefore nondecreasing with \(\mu\) at fixed contact geometry.
Lower-Tail Ranking¶
An approach command specifies palm travel direction, roll, preshape, and closing coordinates. In pose-perturbation trials, the planner adds a translation in the object frame and left-multiplies the palm rotation by a rotation from a sampled rotation vector. It applies the same rotation to the travel direction, repeats the approach and closure, then scores the resulting contacts over the friction grid. Failed approaches or closures contribute zero quality.
Let \(Z\) be a reward with larger values indicating greater wrench capacity. For tail mass \(\beta\in(0,1]\), the lower-tail conditional value-at-risk is
The planner treats all pose–friction pairs, including the nominal pose row, as equally weighted samples. Write \(z_{(1)}\le\cdots\le z_{(M)}\), \(m=\lfloor\beta M\rfloor\), and \(r=\beta M-m\). The empirical score is the sum of the \(m\) smallest scores plus \(r z_{(m+1)}\), divided by \(\beta M\). For example, the lowest \(37.5\%\) of four equally weighted scores \(0,0.03,0.10,0.11\) has average \((0+0.5\cdot 0.03)/1.5=0.01\). The boundary term is omitted at \(r=0\). A score of zero results when at least \(\beta\) of the mass has zero quality. At \(\beta=1\) the ranking uses the sample mean. With one pose and one friction coefficient, it uses the nominal score at the specified coefficient. For the discrete probability treatment, see Rockafellar and Uryasev.
FIRMGrasp evaluates the force-closure margin at the mean of the adverse friction tail. In the planner above, the tail is taken over the quality scores obtained after simulating closure and evaluating the friction grid. In task mode, each score is the supported load multiplier \(Q\).
The planning guide maps these quantities to configuration fields and candidate records.