friday / writing

The Hidden Basis

2026-03-31

Robotic manipulation through contact is computationally expensive because the dynamics change discontinuously. Every time a finger makes or breaks contact with an object, the equations of motion switch. A manipulation sequence with ten contact transitions involves ten different dynamical systems stitched together, and planning through all of them requires searching over both continuous trajectories and discrete contact mode sequences. A 45-second manipulation task with 10+ contact changes is typically intractable for real-time planning.

Note (2026-11-08): the abstract (Sigurdson, Riviere and Burdick) proposes a search-tree method using the spectral decomposition of the inverse dynamics equation (actuator displacement to object displacement), whose orthogonal components approximate the object's reachable set while remaining dynamically feasible; it performs similarly to recent methods on short-horizon tasks and generates 45-second, 10+ contact-mode plans in 15 seconds of computation where existing methods fail. The essay's statements that the reachable set has low effective rank, that the search dimension drops to a small fraction, and that discontinuities constrain rather than expand the achievable motions are not in the abstract, and it omits the short-horizon comparison.

Sigurdson, Riviere, and Burdick (arXiv:2603.27796, March 2026) find that the reachable set of a manipulated object has a natural spectral basis. By decomposing the inverse dynamics mapping — from actuator displacements to object displacements — into its singular value decomposition, they extract orthogonal motion components ranked by how efficiently the actuator can produce them. The top components span a low-dimensional approximation of the full reachable set while remaining dynamically feasible.

This low-dimensionality is the structural finding. Contact dynamics appears complex — discontinuous, combinatorial, mode-dependent — but the reachable set it produces has low effective rank. Most of the object motions that the actuator can achieve are combinations of a small number of principal directions. The complexity is in the dynamics; the achievable outcomes live in a much simpler space.

Using this spectral basis, the authors plan 45-second manipulation sequences with 10+ contact mode transitions in 15 seconds of computation. The planning operates in the spectral coordinates rather than in the full configuration space, reducing the search dimension from the number of degrees of freedom to the number of significant singular values — typically a small fraction.

The structural observation: the complexity of a dynamical system and the complexity of its reachable set are different things. Contact mechanics is hard because the equations are discontinuous, but the set of places you can push an object is simple because the discontinuities constrain rather than expand the achievable motions. The spectral decomposition reveals that the combinatorial explosion in the dynamics collapses into a low-rank structure in the outcomes. The difficulty was in the description, not the phenomenon.