Introduction

[1] Tonneau S., Pettré J., Multon F.: Task EfficientContact Configurations for Arbitrary Virtual Creatures.In Proceedings of Graphics Interface 2014, pp. 9-16.[2] Tan J., Liu C. K., Turk G.: Stable Proportional-Derivative Controllers. IEEE Computer Graphics andApplications, 31(4):34–44, July 2011.

We present our ongoing work to create realisticclimbing animations for articulated 3D characters. Ourgoal is to automatically synthesize physically plausiblemotions for highly detailed models. Specifically, forcharacters with attached hands, giving them truegrasping capabilities. Our method chooses limbconfigurations that are suitable for climbing, favoringpostures with the ability to effectively exchange forcesbetween the character and its environment. Graspingpostures are learned that support the character as itclimbs in a dynamic environment. And a novel linearfeedback controller is used to adjust the grip strengthin order to deal with force variations at the wrist.

The physics-based character model has 83 degrees offreedom (DOF), including 21 DOF for each hand, and amass distribution similar to that of a real human. TheRTQL8 physics engine is used for real-time simulation.

Discussion & Future Work

The preliminary results shown here are encouraging.we are confident that realistic climbing animations forhigh fidelity character models may be synthesizedwith our approach. Aside from an offline optimizationstep, little computational overhead is required,making it suitable for games and other real-timeapplications. Currently, we are investigating methods

A heuristic based on the force transmission ratio isused to determine whether any limbs should berepositioned. The FTR is calculated by mapping thepotential joint torques to a set of potential forces atthe end effector, visualized by an ellipsoid in theabove figure. If joints of an active limb are near limits,or if the FTR in the climbing direction drops below acertain threshold, a new target handhold is selectedfor that limb. A kinematic sampling process based onthe method of Tonneau et al. [1] is used to select areachable handhold with a skeleton configuration thatbest allows the character to push / pull in therequired direction with the limb being retargeted.

Character Model

Controllers & Strategies

Tim Gorm Kaas-Rasmussen Olsen1,2, Sheldon Andrews1, and Paul G. Kry1

McGill University1 University of Copenhagen2

Computational climbing for physics-based characters

Motion Planning

Grasp Optimization

Wrist Inverse Dynamics

Grasp SPD

Reach for handhold

Close handPre-shape

Wrist track spline

Grasping

Grasp SPD

Advance climbing

Body SPD – Non-ragdoll effects, track kinematic pose

Adjust Grip

Finger JTF

Motion planning

Limb JTF

Pull/push to advance

Phases

Controllers

States

References

A state machine is used to activate controllers depending on the phase ofmotion and state of the simulation. Various strategies are used to accomplishthe climbing task.

Body

Limb

Hand

Inverse dynamics is used to re-target alimb not currently supporting the character.The end effector body (palm) is maneuveredto an orientation and position for pre-shaping and grasping the target handhold.The trajectory is stored as a Hermite spline.

A virtual force is generated at each supporting limbaccording to the desired climbing direction 𝑣. Limb joint

torques are computed by 𝜏 = 𝐽𝑙𝑇 𝑣 , where 𝐽𝑙 is the linear

Jacobian of the end effector for limb 𝑙.

A pose tracking controllermaintains a natural configurationfor joints not actively used forclimbing. SPD servos [2] are usedhere too.

The grip controller provides online adjustments to the grasp bymodulating contact forces at each finger. The force needed tosupport the character 𝑓𝑠 is estimated from inertial and externalforces. A linear feedback rule is used to compute the forceadjustments at each finger segment:

𝑓𝑖 = 𝛼 𝑛 𝑛𝑇𝑓𝑠 .Here, 𝑛 is the unit length contact normal at the finger segment and

𝛼 is a gain parameter. Joint torques are computed by 𝜏 = 𝐽𝑖𝑇𝑓𝑖,

where Ji is the Jacobian of finger segment 𝑖 at the contact point.

The grasp controller pre-shapes and closes the hand according to agrasping posture learned by the optimization step. Torques at fingerjoints are computed using stable proportional derivative (SPD) servos [2].

Example videos goo.gl/HNrrcz

to perform online graspplanning, eliminating theneed for a costly non-linearoptimization step. We hopeto make climbing aubiquitous activity availableto characters in any physics-based virtual environment.Furthermore, data collectedwith the sensor augmentedclimbing wall, shown in thefigure on the left, will beused to determine otherclimbing strategies.

An offline CMA optimization step learns graspingpostures that maximize the forces exchanged due tocontacts between the fingers and a candidatehandhold. A database of grasping postures is learnedfor each handhold, and for various pulling directions.Rather than considering all 21 DOF of the handindependently, the optimization uses a synergy ofhand postures captured from real climbing tasks by aShapeHand sensor. This lowers the dimensionality ofthe search to 3 DOF for the reduced pose basis and 6DOF for the position and orientation of the palm.