Contact-rich assembly is where robot learning demos go to die. A peg that is 0.4 mm too tight, a bottle that cracks above a few newtons — these tasks punish the brute-force policies that ace benchmarks built on picking and placing. A paper submitted to arXiv on July 23 by Kyupaeck Jeff Rah and Midum Oh proposes a different division of labor: FORGE-plus, a two-layer framework that passed all 256 of its evaluation episodes on fragile-object placement and tight-clearance gear insertion without breaking a single workpiece.
An LLM That Budgets Force Instead of Producing It
The core insight is that large language models are poor at emitting continuous force commands but remarkably good at reasoning about limits. FORGE-plus keeps its LLM frozen and text-only: given a task description, the model assigns each object a per-contact force ceiling and, when things go wrong, selects recovery maneuvers from a fixed menu — guided by textual descriptions of force signatures rather than raw sensor streams. Low-level execution stays with conventional force controllers; the language model never touches the torque loop. The result is a system whose safety envelope is explicit, auditable and set before the arm moves.
5.4 N Average Peak Force, Zero Breakage
Evaluated on fragile bottle placement and gear insertion with a 0.4 mm clearance — using a Robotiq 2F-140 gripper on a Franka Panda arm — FORGE-plus averaged a peak contact force of just 5.4 N across the benchmark. When the researchers deliberately injected slip mid-episode, the textual-signature recovery step rescued between 40% and 64% of trials depending on the condition, turning would-be failures into completed insertions.
The Negative Results Matter as Much
Equally valuable is what the paper reports does not work. PPO-trained policies fail under strict force constraints, and learned release strategies underperform their scripted counterparts — evidence that for fragile assembly, careful decomposition beats end-to-end learning when hardware margins are measured in newtons.
The caveat is stated just as clearly: every evaluation ran in pure rigid-body simulation, with no sim-to-real transfer claims. But the recipe is cheap to adopt — a frozen model, a force budget, a recovery menu — and it gives integrators a template for putting language-model reasoning onto real lines without handing it the servo loop.
