Contact IK

Inverse kinematics that lets the whole arm respond to touch.

Differentiable IK · Whole-arm contact · Isaac Lab

Reach with the hand.
Respond with the whole arm.

A touch on the arm changes the problem. Contact IK puts that information inside the inverse-kinematics objective, so the robot can adjust its posture while pursuing a hand goal.

Watch it happen

Same goal. A different solve.

Offline IK visualizations

Pose-only IK and Contact IK start from the same recorded state, with the same hand goal and joint bounds. Cyan marks the goal; amber marks the fixed contact input.

Visual description

Both arms start in the same configuration. The cyan marker is the desired hand pose. The amber point is a recorded contact location, and its halo shows the contact penalty region. The left solve optimizes the pose objective and joint bounds. The right solve adds the contact objective. As the optimizer progresses, the postures diverge. The readouts retain the remaining hand error and contact cost.

Cyan: hand goalAmber: contact inputReadouts: remaining task error & contact cost

The idea

Give contact a place in the objective.

01

Locate the touch

Express an observed contact point in the robot's world frame. Link geometry tells the solver how sampled surface points move with each joint.

02

Combine the objectives

Keep the hand-pose goal, add a penalty for sampled points near the contact input, and include joint-limit terms.

03

Update the posture

Differentiate through forward kinematics and update the joints. The contact term can change the solution across the arm.

Find a joint configuration
task error+λ · contact cost+joint-limit cost

The comparison changes only the contact weight: λ = 0 on the left, λ = 1 on the right.

A closer look at the contact term

The illustrated formulation sums squared penalties for sampled points within an 8 cm radius of a fixed observed contact point. The samples form a synthetic geometric atlas. The cost is an optimization signal; it is not a pressure measurement or a guarantee of collision-free motion.

Both rendered conditions project each optimizer update to the same joint bounds. This explicit visualization modification keeps the displayed configurations bounded. The original study below used the unprojected formulation with a final command-boundary projection.

Explore the geometry

Look closer.

Contact geometry, from every angle

An orbit view makes the contact input and the changing arm posture easier to see. The amber halo illustrates the formulation's 8 cm penalty radius around a fixed observed point.

Visual description

The camera moves around a Franka Panda arm while Contact IK updates its joint configuration. The cyan marker remains at the hand goal. The amber contact input remains fixed in world coordinates. Its halo represents an objective region, rather than measured pressure or a physical obstacle boundary.

Inside the solve

Follow the optimizer iteration and remaining hand error. These are retimed solver iterates from a joint-bounded visualization variant of the original objective.

Visual description

Contact IK combines a hand-pose objective, a contact penalty, and joint-limit terms. Differentiation through the robot's kinematics gives a joint-space update. The displayed iteration advances from zero to one thousand. The remaining hand error stays visible so the trade-off can be inspected.

Rendered in NVIDIA Isaac Lab with the Franka Panda asset from Isaac Sim. The clips replay and retime offline optimizer iterates from recorded exploratory states. They show a joint-bounded variant of the original Contact IK objective; closed-loop force control and hardware performance require separate evaluation.

Results & context

What the original study showed

The original formulation study evaluated 28 recorded Franka states, including 12 with contact. The contact term reduced the avoidance residual in every contact-positive case. Pose accuracy, solve speed, and reliable convergence remained limiting.

12 / 12Contact cases with a lower avoidance residual
0 / 12Contact cases meeting the full pose-success criterion
28 statesKinematic evaluation; physical validation pending
Original ContactIK avoidance residual before and after optimization for twelve contact-positive cases
Original fixed-weight formulation: lower avoidance residual in all 12 contact-positive cases, with a nonzero residual in 11. These historical numbers belong to the original evaluation, rather than the joint-bounded render variant.

Where the research goes next

My broader work connects contact sensing to manipulation along the arm and body. The first hardware direction is the Spot arm with the existing force-sensing skins, using bounded, supervised experiments. Related simulation studies explore pushing, bracing, and pivoting.

ContactIK-MPC is a separate research extension. It adds learned objectives and short-horizon contact prediction. The videos on this page demonstrate the original geometric idea and its joint-bounded visualization variant.

Related whole-body simulation

A planar pushing and bracing setup helped diagnose whole-body contact tasks. It is a separate simulation study, with learned-policy and hardware validation still open.

Planar robot and funnel used in a separate contact-manipulation simulation

Rendering tools: NVIDIA Isaac Lab · Isaac Sim robot assets · Media provenance. Clips are silent and include optional English descriptions and MP4 downloads.