Skip to content
AI-generated research idea🟠Medium🔬Significant ResearchRelated work found

Automated Synthesis of Force Primitives and Physics Estimators via Code Generation

Generate parameterized admittance control primitives and trace translation estimators using LLM code generation to handle contact geometries beyond hand-crafted tool menus.

Time

several weeks

Compute level

Single GPU or less

Resources

MuJoCo/robosuite simulation environment, OpenAI/Anthropic API access, and a standard GPU workstation.

What the paper used

Franka FR3 arm with Polymetis, robosuite 1.5.1 on MuJoCo 3.2.7, Isaac Lab 3.0.0b2, and remote frontier APIs (GPT-6 Astra, GPT-6.1 Sol, Claude Fable 5.2).

Prerequisites

Admittance and impedance control, Python code generation with LLMs, robotic simulation in MuJoCo.

Scope

modify existing code

The Gap

Section 3.3 and Section 3.5 define a fixed library of 12 hand-engineered force primitives (e.g., search, slide, wiggle, screw) and five hardcoded physics estimators (Equations 4, 5, and B1-B4). When encountering novel interaction mechanisms or geometry that do not match these predefined skills, the system cannot adapt its control laws or extract relevant physical quantities.

“Both the primitive menu and the physics estimators are specified by hand, and each task relies on a hand-written instruction, leaving automatic primitive synthesis and open-ended task specification to future work.”
From the paper, Section J.1. Matched against the PDF text.

Limitation

The primitive menu and physics translation estimators are manually specified, restricting zero-shot execution to contact modes anticipated by the framework designers.

Proposed Solution

Implement an in-context code-generation module where the frontier model, prompted with high-level assembly requirements and low-level safety contracts, synthesizes custom Python admittance laws and event predicates. Pair this with an automated translation synthesizer that fits mathematical models to logged force-displacement traces (such as spring stiffnesses, damping profiles, or thread pitch) using symbolic regression, returning structured physical measurements to the agent loop.

The paper’s authors list this direction as future work, so they may already be pursuing it.

Why This Matters

Allowing autonomous primitive synthesis would enable zero-shot frontier models to tackle irregular assemblies, flexible materials, and specialized industrial fixtures without human intervention.

How to Get Started

How to test it: Evaluate on 5 novel contact-rich assembly tasks in HiddenPhys with unseen geometric constraints against the fixed 12-primitive ForceHarness baseline, measuring task completion rate. If the model-generated primitives fail to execute safely within the 25 N safety boundary on over 20% of seeds, stop and re-evaluate the prompt contracts.

  1. Inspect the open-source ForceHarness repository and isolate the primitive control loop structure in RUN (Figure A1).
  2. Implement a code generation wrapper that prompts GPT-6 or Claude to emit Python control laws adhering to the `law(F, p, t)` interface.
  3. Evaluate the synthesized primitives in the robosuite Panda simulation on a non-standard peg-in-hole task with complex cross-sections.

Suggested by AI. Adapt it to your setup.

Has This Been Done?

Related work found

Works like GenCHiP and FORCA explore code generation for contact-rich robot manipulation, while Act-Observe-Rewrite and RoboCodeX examine multimodal coding agents. However, these listed works do not explicitly cover the combination of in-context code generation for custom admittance laws combined with automated symbolic regression for physics parameter fitting from force-displacement traces.

We searched OpenAlex and arXiv for “LLM code generation contact primitives compliant manipulation” and “Automated Synthesis of Force Primitives and Physics Estimators via Code Generation” and an AI compared the idea with the top 10 results. A small search can miss relevant work.

Source Paper

ForceHarness: Turning Force into Action and Physics for Frozen Frontier Models

Ke Sun, Kevin Yuchen Ma, Yike Zhang +2 more · 2026

ForceHarness enables API-only, frozen frontier multimodal models to perform contact-rich robotic manipulation without demonstrations or fine-tuning. It introduces a three-layer architecture comprising force sensing and calibration, self-terminating closed-loop force primitives, and a physics translation layer that converts force traces into interpretable physical quantities. Across real-robot hardware and simulation benchmarks including HiddenPhys and ForceBench, the system dramatically improves zero-shot manipulation success and matches expert force controllers.

Related Ideas

Is this a good idea?

Share onXLinkedIn