Constraints and Planning for Forceful Robotic Manipulation (Rachel Holladay, MIT)

In this talk

1. 3:23

   Model force limits in tool use

   Holladay frames a fixed sequence of manipulation steps as a constraint satisfaction problem over grasps, poses, and motion paths. For a hammer pulling a nail, the robot must choose a grasp and path that keep every joint in the forceful kinematic chain stable. She checks robot joint torque limits and whether friction at the grasp can withstand the wrench; the same approach is shown for screwdriving, cutting, and tightening a bolt.

2. 24:24

   Let the robot choose action sequences

   A fixed sequence limits adaptation when the environment changes, so Holladay represents actions as operators with preconditions, effects, and parameters. The robot can combine operators such as picking, placing, fixturing, and applying a push-twist to construct and evaluate its own constraint graph. For a childproof bottle, possible strategies include relying on table friction, moving the bottle to a vise or rubber mat, or using a tool.

3. 31:03

   Plan with compliant control

   Holladay describes Cartesian impedance control, which relates motion and force like a mass-spring-damper. The planner selects set points and stiffness values to apply the bottle cap’s pushing and twisting forces, while checking that the relevant contact chains remain stable. PDDLStream combines samplers for continuous choices, such as grasps and paths, with a task planner that searches candidate action sequences.

4. 40:56

   Account for uncertainty and limits

   The framework is also explored for cutting vegetables, with cutting simplified into a downward press followed by a horizontal slice. Holladay notes that impedance control can stall on rigid objects, and that ignoring forces from the cut can lead to a grasp that slips. To favor reliable plans, the team perturbs estimates such as friction and contact area, estimates success rates, and assigns higher cost to actions that fail more often.

5. 47:04

   Identify next challenges

   Holladay describes force amplification through tools, planning under uncertainty about the robot’s state and future outcomes, and manipulation of elastic objects as ongoing directions. Elastic tasks may require maintaining contact while changing which surfaces support the force. In questions, she adds that the current action space limits plans to combinations of specified operators, and that the methods assume models of objects and physical properties.