Research map / Reinforcement learning / Visual and representation RL
SRSA: Skill Retrieval and Adaptation for Robotic Assembly Tasks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SRSA, which retrieves a skill from a library of assembly policies by predicting zero-shot transfer success from geometry, dynamics, and expert-action features, then fine-tunes it with PPO plus self-imitation learning.
Problem
Data-efficient learning of new contact-rich robotic assembly tasks; the challenge is choosing which prior specialist policy from a skill library to fine-tune on a new task.
Approach
Predict zero-shot transfer success of every library policy on a new task using a learned predictor over jointly learned geometry (PointNet autoencoder), forward-dynamics, and expert-action (from disassembly paths) task embeddings; retrieve top-k by prediction, pick best by zero-shot success, then fine-tune with PPO and self-imitation learning.
Key results
- In the dense-reward setting SRSA reaches 82.6% average success on 10 test tasks vs 69.4% for AutoMate (19% relative improvement) with 2.6x lower std across seeds.
- SRSA needs at least 2.4x fewer training epochs than AutoMate to reach target success thresholds on test tasks.
- Policies fine-tuned in simulation achieve a 90% mean success rate when deployed zero-shot on real Franka robot assembly tasks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| retrieve a specialist skill and fine-tune it with self-imitation learning instead of training from scratch | AutoMate specialist policy learning · AutoMate test tasks, dense reward | success rate: 69.4% -> 82.6%; 19% | Helped |
| retrieve a specialist skill and use self-imitation learning instead of training from scratch | AutoMate specialist policy learning · AutoMate test tasks, sparse reward | success rate: 135% | Helped |
| retrieve a specialist skill and fine-tune it instead of training from scratch | AutoMate specialist policy learning · AutoMate test tasks, dense reward | training epochs to reach a success threshold: at least 2.4 times fewer training epochs | Helped |
| retrieve and fine-tune specialist policies while expanding the skill library | AutoMate continual learning · 90 tasks outside the initial skill library | sample efficiency: 84% | Helped |
| use SRSA policy fine-tuning | AutoMate continual learning · 100 tasks | average success rate: 70% -> 79% | Helped |
| use SRSA retrieval instead of baseline retrieval approaches | AutoMate skill retrieval · 10 test tasks, zero-shot transfer | success rate: around 20% higher | Helped |
| use the retrieved specialist policy rather than a generalist policy for initialization (ablation) | SRSA policy fine-tuning · AutoMate unseen tasks | success rate | Helped |
| retrieve skills using geometry, dynamics, and expert-action features rather than geometry embeddings alone (ablation) | SRSA policy fine-tuning · AutoMate test tasks | learning efficiency | Helped |
| add self-imitation learning to fine-tuning (ablation) | SRSA policy fine-tuning · AutoMate test tasks | learning stability | Helped |
| add self-imitation learning to fine-tuning (ablation) | SRSA policy fine-tuning · AutoMate test tasks | standard deviation of success rate across runs | Helped |
Surprising findings
- Fine-tuning a generalist policy (SRSA-Gen) gave weaker initialization and lower asymptotic performance than fine-tuning a retrieved specialist policy.
Limitations
- Does not address assemblies requiring rotational or helical motion such as nut-and-bolt insertion.
- Focuses on specialist (single-task) policies rather than generalist policies.
- Real-world success (90%) falls short of the 99+% needed for industrial deployment.
Experimental setup
Datasets: AutoMate benchmark (100 two-part assembly tasks, meshes, disassembly trajectories)
Benchmarks: AutoMate benchmark
Metrics: zero-shot transfer success rate success rate relative improvement standard deviation across seeds training epochs to reach success threshold real-world success rate
Baselines: AutoMate Signature retrieval Behavior retrieval (VAE) Forward dynamics retrieval Geometry retrieval (PointNet) SRSA-Geom SRSA-noSIL SRSA-Gen
Builds on: PPO self-imitation learning AutoMate context-based meta-RL
Closest papers by meaning
- SAM2Act: Integrating Visual Foundation Model with A Memory Architecture for Robotic Manipulation (ICML 2025 · similarity 0.88)
- RobotSmith: Generative Robotic Tool Design for Acquisition of Complex Manipulation Skills (NeurIPS 2025 · similarity 0.86)
- Robots Pre-train Robots: Manipulation-Centric Robotic Representation from Large-Scale Robot Datasets (ICLR 2025 · similarity 0.86)
- Predictive Inverse Dynamics Models are Scalable Learners for Robotic Manipulation (ICLR 2025 · similarity 0.86)
- AHA: A Vision-Language-Model for Detecting and Reasoning Over Failures in Robotic Manipulation (ICLR 2025 · similarity 0.86)
- Rapid Motor Adaptation for Robotic Manipulator Arms (CVPR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.