Hugging Face Releases ENCORE: Few-Shot Agentic Discovery of Manipulation Strategies
Summary:Hugging Face introduces ENCORE, a novel framework that enables agents to discover manipulation strategies through a few demonstrations. The framework decomposes demonstrations into multi-view keyframes, gripper events, frame strips, and full trajectories, allowing agents to learn and write policy programs. Experiments on LIBERO-PRO and RoboDojo tasks show that ENCORE outperforms existing methods, particularly in complex tasks. Additionally, ENCORE has been successfully applied to a real bimanual
Technical Mechanism Analysis
The core mechanisms of the ENCORE framework include the following steps:
- Demonstration Decomposition: Decompose a few demonstrations into multi-view keyframes, gripper events, frame strips, and full trajectories. These decomposed data serve as evidence for the agent to learn from, rather than traditional training data.
- Policy Program Writing: The agent writes a policy program against a fixed perception and action API by studying the decomposed data.
- Iterative Optimization: The agent iteratively optimizes the policy program over a few development rollouts.
- Sealed Evaluation: The agent runs the policy program in a sealed evaluation where the success signal is never revealed, ensuring the objectivity of the evaluation.
Engineering Trade-offs and Performance
ENCORE demonstrates strong performance on LIBERO-PRO tasks, with the first program succeeding in 50% of the perturbed tasks and one task without demonstrations. In contrast, the strongest prior agentic system achieves only 89.3% success with the same language model. Additionally, on RoboDojo tasks, ENCORE fails to succeed without demonstrations, highlighting the importance of demonstrations in complex tasks.
The application of ENCORE on a real bimanual robot further validates its practicality. By learning from five demonstrations, ENCORE successfully completes cube handover and cup inversion tasks, demonstrating its potential in practical scenarios.
Developer Implementation and Deployment Recommendations
For developers, ENCORE offers a new paradigm for agentic learning, enabling the discovery of complex task strategies through a few demonstrations. Here are some deployment recommendations:
- Data Preparation: Ensure the quality and diversity of demonstration data to support effective learning by the agent.
- API Design: Design clear perception and action APIs to allow the agent to accurately understand and execute tasks.
- Evaluation Mechanism: Establish an objective evaluation mechanism to prevent the leakage of success signals during evaluation.
- Continuous Optimization: Continuously optimize the agent's policy program based on feedback from practical application scenarios.
Conclusion
The ENCORE framework provides an efficient method for discovering agentic manipulation strategies, particularly excelling in complex tasks. Its successful application on a real robot demonstrates its significant potential in practical scenarios.
— END —Source: Hugging Face Daily Papers (2026-09-29)
Tags: #Hugging Face #Agentic Systems #Machine Learning #Robotics #Strategy Discovery
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments