Research map / Reinforcement learning / Visual and representation RL

Compliant Residual DAgger: Improving Real-World Contact-Rich Manipulation with Human Corrections

NeurIPS 2025 · Reinforcement learning › Visual and representation RL · approach: Dexterous manipulation transfer · 5 citations (OpenAlex)

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 CR-DAgger, a compliant kinesthetic intervention interface for on-policy delta correction plus a force-aware residual policy trained on correction data, improving diffusion base policies on four contact-rich real-world tasks.

Problem

Improving pretrained robot manipulation policies for contact-rich tasks with a small amount of human correction data, where existing takeover-based interventions cause discontinuities and retraining/finetuning are costly or unstable.

Approach

An admittance-controlled kinesthetic handle lets humans apply delta corrections while the base policy keeps running, with velocity tracking and look-ahead to reduce tracking-error-induced intention misinterpretation; a lightweight residual policy predicts residual pose deltas and target wrenches from images, proprioception and force.

Key results

• Compliant Residual policy with on-policy delta data improves base policy success by 60% on book flipping and 50% on belt assembly; the paper's headline claim is a 64% average improvement across four tasks.
• Compliant Residual reaches 100% on the book flipping final stage, while the position-only Residual drops from 70% to 35% at that stage.
• Retraining from scratch gives insignificant gains (1.67% average drop on book task) and finetuning underperforms the base policy by 30% on the book task.

Measured effects reported in the paper

Surprising findings

• Finetuning the base policy on correction data often underperforms the original base policy (30% average drop on book flipping).
• More correction batches do not monotonically improve performance; three batches performed slightly worse than two on belt assembly.

Limitations

• Base policy needs a reasonable success rate (~10-20% minimum) for the residual to learn effectively.
• Limited robustness to large variations since only a small amount of correction data is used.
• Kinesthetic teaching requires more labor than teleoperation and is limited to single-arm setups.

Experimental setup

Datasets: real-world robot demonstrations for base policy (150-405 per task) on-policy human correction episodes (50-100 per task)

Benchmarks: Book flipping (real world) Belt assembly (real world, NIST board) Cable routing (real world) Gear insertion (real world, NIST board)

Metrics: success rate velocity magnitude fingertip trajectory distribution

Baselines: Observe-then-Collect Take-over-Correction (HG-DAgger) Retrain Policy Finetune Policy Finetune with KL Regularization Residual Policy (position-only) Base policy without force

Builds on: DAgger HG-DAgger Diffusion Policy admittance control

Closest papers by meaning

• DexFlyWheel: A Scalable and Self-improving Data Generation Framework for Dexterous Manipulation (NeurIPS 2025 · similarity 0.87)
• ManipTrans: Efficient Dexterous Bimanual Manipulation Transfer via Residual Learning (CVPR 2025 · similarity 0.87)
• ForceVLA: Enhancing VLA Models with a Force-aware MoE for Contact-rich Manipulation (NeurIPS 2025 · similarity 0.85)
• Phoenix: A Motion-based Self-Reflection Framework for Fine-grained Robotic Action Correction (CVPR 2025 · similarity 0.85)
• SRSA: Skill Retrieval and Adaptation for Robotic Assembly Tasks (ICLR 2025 · similarity 0.84)
• Efficient Residual Learning with Mixture-of-Experts for Universal Dexterous Grasping (ICLR 2025 · similarity 0.84)