New Algorithm Lets Legged Robots Pick Smarter Support Contacts

Robots that combine legs and an arm—so-called loco-manipulators—often need to brace part of their body against a wall, ledge, or other surface to gain stability while reaching or pushing on something. But choosing the right surface to lean on isn't simple: a contact that offers strong physical support can also block the very motion the robot needs to complete its task. A new paper, as reported by arXiv (cs.RO), frames this as a joint optimization problem and introduces a method to solve it efficiently.

The researchers describe the challenge through three "capability" measures that a robot retains after committing to a support contact: residual wrench (how much extra force or torque it can still exert), end-effector reach (how far its arm can extend), and base mobility (how freely its body can still move). These capabilities are weighed against the cost of acquiring a given contact in the first place. Evaluating every possible contact option this way is expensive, because each one normally requires running a full whole-body optimization to see how the robot would perform.

To cut that computational burden, the team built Capability-Tradeoff Contact Selection (CTCS). Instead of exactly solving the whole-body optimization for every candidate contact, CTCS first screens out contacts and configurations that fail basic feasibility checks. It then clusters similar candidates on each surface together and estimates their capabilities using local sensitivity analysis anchored to a small number of exact evaluations. The system spot-checks these estimates with additional exact evaluations where needed, ranks the remaining candidates by their predicted capability trade-offs, and finally runs full, exact optimization only on a short list of top contenders to make the final selection.

The team validated CTCS in both simulation and hardware experiments using a Unitree Go2 quadruped equipped with an AgileX NERO arm, across 392 distinct task conditions and nine available support surfaces. According to the results, CTCS outperformed both ground-only support (where the robot never braces against anything) and fixed-contact strategies (where a single predetermined contact is always used), because it could adaptively pick surfaces offering the best trade-off for each specific task. Compared with exhaustively evaluating every candidate contact exactly, CTCS delivered roughly a 3x speedup while closely matching the mean objective value that exhaustive search achieved.

The work points toward more capable field and service robots that can autonomously decide not just how to move, but what to lean on while doing so—a small but practical piece of the puzzle for robots operating in cluttered, real-world environments like construction sites, disaster zones, or warehouses where improvised bracing points are common. The paper, titled "Contact as a Decision Variable: Capability-Tradeoff Contact Selection for Legged Loco-Manipulation," spans nine pages with six figures.