[Paper] Contact-Anchored Policies: Contact Conditioning Creates Strong Robot Utility Models

Source: arXiv - 2602.09017v1

Overview

The paper introduces Contact‑Anchored Policies (CAP), a new way to teach robots to manipulate objects by conditioning policies on where the robot makes contact rather than on abstract language commands. By treating each contact point as a modular “utility model,” the authors can rapidly prototype and debug in a lightweight simulator (EgoGym) before deploying to real hardware, achieving strong generalization with only a few dozen hours of demonstrations.

Key Contributions

• Contact‑based conditioning: Replaces language prompts with explicit 3‑D contact points, giving the robot a concrete physical reference for action planning.
• Modular utility‑model library: Decomposes a monolithic policy into reusable sub‑models that predict the utility of a contact configuration, enabling easier debugging and transfer.
• Real‑to‑sim iteration loop: Introduces EgoGym, a fast‑to‑run simulation benchmark that mirrors the real‑world setup, allowing rapid identification of failure modes and data augmentation.
• Data efficiency: Demonstrates strong performance on three core manipulation skills using only 23 h of human‑provided demonstrations.
• Zero‑shot superiority: Outperforms large vision‑language agents (VLAs) by ≈56 % in zero‑shot tests on unseen environments and robot embodiments.
• Open‑source release: All code, simulation assets, hardware designs, and datasets will be publicly available, lowering the barrier for reproducible robot learning research.

Methodology

1. Contact Representation:

   • Each demonstration is annotated with a set of 3‑D points where the robot’s end‑effector (or other links) touches the environment or object.
   • These points are fed to a neural utility model that predicts the expected “task success” for that contact configuration.

2. Utility Model Library:

   • Instead of a single end‑to‑end policy, CAP builds a collection of lightweight models (one per skill or contact primitive).
   • At runtime, the system selects and composes the relevant utilities to generate a full action sequence.

3. EgoGym Simulation Loop:

   • A stripped‑down physics simulator that mirrors the real robot’s kinematics and sensor suite.
   • Researchers run large‑scale sweeps of contact configurations, automatically flagging those that lead to failures (e.g., slippage, unreachable poses).
   • Failure cases are fed back into the data collection pipeline, either by augmenting the simulation or by guiding additional real‑world demos.

4. Training & Deployment:

   • The utility models are trained with supervised learning on the 23 h of demonstrations, using a simple binary success label.
   • During deployment, a planner samples candidate contacts, queries the utility models, and executes the highest‑scoring plan on the physical robot.

Results & Findings

• Generalization: CAP transferred to new robot arms (different kinematic chains) and novel tabletop layouts without any fine‑tuning.
• Sample Efficiency: The same performance level would require >200 h of data for language‑conditioned baselines.
• Simulation‑Real Gap: The EgoGym loop reduced the sim‑to‑real discrepancy to <5 % in success rates, a dramatic improvement over naïve sim‑only training.

Practical Implications

• Faster Prototyping: Engineers can iterate on manipulation pipelines in seconds using EgoGym, dramatically cutting down hardware testing cycles.
• Robust Deployment: By anchoring policies to physical contacts, robots become less prone to misinterpretations of ambiguous language commands, leading to safer operation in unstructured environments (e.g., warehouses, homes).
• Modular Skill Libraries: Companies can build a catalog of reusable contact utilities (grasp, push, slide) that can be mixed‑and‑matched for new tasks, reducing the need for task‑specific retraining.
• Lower Data Costs: Small‑scale data collection (a few dozen hours) is sufficient, making robot learning feasible for startups and research labs without massive data‑labeling budgets.
• Hardware‑agnostic Solutions: Because contact points are expressed in world coordinates, the same utility models can be deployed on different robot platforms with minimal calibration.

Limitations & Future Work

• Contact Annotation Overhead: The current pipeline still requires manual labeling of contact points in demonstrations, which may not scale to highly complex tasks.
• Limited Skill Set: The study focuses on three fundamental manipulation primitives; extending CAP to high‑dimensional tasks like assembly or deformable‑object handling remains open.
• Simulation Fidelity: While EgoGym is lightweight, it abstracts away fine‑grained dynamics (e.g., friction variations) that could affect performance on highly tactile tasks.
• Real‑World Perception: The approach assumes accurate 3‑D perception of contact locations; noisy depth sensors could degrade utility predictions.

Future research directions include automated contact extraction from raw video, expanding the utility library to cover a broader taxonomy of contacts, and integrating tactile feedback to refine contact‑conditioned policies on‑the‑fly.

Authors

• Zichen Jeff Cui
• Omar Rayyan
• Haritheja Etukuru
• Bowen Tan
• Zavier Andrianarivo
• Zicheng Teng
• Yihang Zhou
• Krish Mehta
• Nicholas Wojno
• Kevin Yuanbo Wu
• Manan H Anjaria
• Ziyuan Wu
• Manrong Mao
• Guangxun Zhang
• Binit Shah
• Yejin Kim
• Soumith Chintala
• Lerrel Pinto
• Nur Muhammad Mahi Shafiullah

Paper Information

• arXiv ID: 2602.09017v1
• Categories: cs.RO, cs.LG
• Published: February 9, 2026
• PDF: Download PDF