Research map / Reinforcement learning / Visual and representation RL
SCaR: Refining Skill Chaining for Long-Horizon Robotic Manipulation via Dual Regularization
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 SCaR, a skill chaining framework that combines adaptive equilibrium scheduling of RL and GAIL rewards for sub-task pre-training with bi-directional adversarial regularization for fine-tuning, improving long-horizon manipulation success in IKEA assembly and kitchen simulation and a real pick-and-place setup.
Problem
Long-horizon robotic manipulation via skill chaining often fails because of imperfect sub-task skill learning and disturbances, which accumulate errors across sequential sub-task skills.
Approach
Dual regularization: AES schedules the balance between environment RL reward and expert-imitation GAIL reward based on imitation progress to strengthen intra-skill dependencies, and a bi-directional discriminator pulls each skill's terminal states toward the next skill's initial states and vice versa to strengthen inter-skill dependencies.
Key results
- SCaR achieves 0.94 average success on six furniture assembly tasks, a 12% improvement over T-STAR (0.85 average).
- On kitchen tasks SCaR reaches 0.78 average success vs 0.58 for T-STAR, an 18% improvement.
- Under perturbation on chair_ingolf, SCaR achieves 0.80 success vs 0.59 for T-STAR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add SCaR dual regularization | skill chaining · Furniture Assembly (All) | task completion progress: 0.85 ± 0.09 -> 0.94 ± 0.03 (12% ↑) | Helped |
| add SCaR dual regularization | skill chaining · Kitchen Organization (All) | task completion progress: 0.58 ± 0.10 -> 0.78± 0.12 (18% ↑) | Helped |
| add SCaR dual regularization | skill chaining · chair_bernhard (perturbed) | success rate: 0.60 ± 0.08 -> 0.85 ± 0.11 | Helped |
| add SCaR dual regularization | skill chaining · chair_ingolf (perturbed) | success rate: 0.59 ± 0.04 -> 0.80 ± 0.13 | Helped |
| add SCaR dual regularization | real-world pick-and-place · 2 sub-tasks | success rate: 70% -> 90% | Helped |
| add SCaR dual regularization | real-world pick-and-place · 3 sub-tasks | success rate: 50% -> 70% | Helped |
| add bi-directional regularization (ablation) | SCaR · chair_ingolf | task completion progress: 0.91± 0.01 -> 0.95± 0.03 | Helped |
| add AES regularization (ablation) | SCaR · chair_ingolf | task completion progress: 0.93± 0.02 -> 0.95± 0.03 | Helped |
| add bi-directional regularization (ablation) | SCaR · chair_ingolf (perturbed) | success rate: 0.63 ± 0.05 -> 0.80 ± 0.13 | Helped |
| add AES regularization (ablation) | SCaR · chair_ingolf (perturbed) | success rate: 0.71 ± 0.07 -> 0.80 ± 0.13 | Helped |
| set the initial RL/IL balance to α = 0.5 (ablation) | SCaR · chair_agne, chair_ingolf, table_dockstra, Extended Kitchen | success rate | Helped |
| increase expert demonstration data from 400 to 800 demos (ablation) | Policy Sequencing, T-STAR, and SCaR · chair_ingolf | success rate: 400 to 800 demos | Helped |
Surprising findings
- Scratch-trained PPO and GAIL achieve near-zero success on the 4- and 5-sub-task tasks, showing the difficulty of long-horizon learning without skill decomposition.
- Imbalanced weighting of the two bi-directional constraints degrades performance, with balanced weighting best.
Limitations
- Sub-task division is predefined and lacks visual or semantic object processing.
- Real-world validation is limited to simple simulated-trained pick-and-place tasks rather than real furniture assembly.
- Performance drops as the number of sub-tasks grows (0.61 on a 6-sub-task kitchen variant).
Experimental setup
Datasets: IKEA furniture assembly environment demonstrations (procedural assembly policy) D4RL Franka Kitchen
Benchmarks: IKEA furniture assembly (chair_agne, chair_bernhard, chair_ingolf, table_lack, toy_table, table_dockstra) Franka Kitchen Extended Kitchen real-world Sagittarius K1 pick-and-place
Metrics: success rate (sub-task completion progress) success rate under perturbation
Baselines: PPO GAIL Fixed-RL-IL SkiMo Policy Sequencing T-STAR
Builds on: T-STAR GAIL ADS (automatic discount scheduling) Policy Sequencing SkiMo
Closest papers by meaning
- ToS: Tree-of-Skill Reinforcement Learning driven by Behavior Tree for Long-Horizon Manipulation (NeurIPS 2026 · similarity 0.88)
- Multi-Stage Manipulation with Demonstration-Augmented Reward, Policy, and World Model Learning (ICML 2025 · similarity 0.87)
- RoboCerebra: A Large-scale Benchmark for Long-horizon Robotic Manipulation Evaluation (NeurIPS 2025 · similarity 0.87)
- VICtoR: Learning Hierarchical Vision-Instruction Correlation Rewards for Long-horizon Manipulation (ICLR 2025 · similarity 0.86)
- Phoenix: A Motion-based Self-Reflection Framework for Fine-grained Robotic Action Correction (CVPR 2025 · similarity 0.86)
- ManiSkill-HAB: A Benchmark for Low-Level Manipulation in Home Rearrangement Tasks (ICLR 2025 · similarity 0.86)
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.