Research map / Reinforcement learning / Imitation learning
CCIL: Continuity-Based Data Augmentation for Corrective Imitation Learning
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 CCIL, which learns a locally Lipschitz-continuous residual dynamics model from expert demonstrations and uses it to generate corrective state-action labels for behavior cloning augmentation.
Problem
Behavior cloning agents are brittle to compounding errors, disturbances, and out-of-distribution states, and existing robustness fixes need interactive experts, extra offline data, or domain invariances.
Approach
Learn a residual dynamics model from expert transitions with a sampling-based local Lipschitz penalty, then solve root-finding problems (BackTrack and DisturbedAction labels) to synthesize corrective state-action pairs near expert support, filtered by rejection sampling, and train BC on the augmented data.
Key results
- On F1tenth racing, CCIL reaches 56.4% success rate vs 31.9% for BC and 39.3% for NoiseBC.
- On the pendulum, the empirical label error of 0.02367 is below the theoretical bound of 0.065.
- CCIL achieves the best result on 10 of 14 tasks according to the authors' summary, and matches BC on the rest.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| augment behavior cloning with CCIL corrective labels | BC · Pendulum | performance | Helped |
| augment behavior cloning with CCIL corrective labels | BC · Discontinuous Pendulum | performance | Helped |
| use a dynamics model without explicitly enforcing Lipschitz continuity to generate labels (ablation) | CCIL · Pendulum | performance | Hurt |
| augment behavior cloning with CCIL corrective labels | BC · F1tenth Racing | success rate: 39.3% -> 56.4% | Helped |
| augment behavior cloning with CCIL corrective labels | BC · F1tenth Racing | average score: 0.62 ± 0.28 -> 0.75 ± 0.25 | Helped |
| augment behavior cloning with CCIL corrective labels | BC · Drone navigation | score | Helped |
| augment behavior cloning with CCIL corrective labels | BC · MuJoCo locomotion and MetaWorld manipulation | return | Mixed |
Surprising findings
- MOREL and MILO fail badly on several tasks (e.g., 0.0% success on F1tenth, and drone returns around -1e8), performing far worse than plain BC.
- Lidar and contact-rich discontinuities hurt model-based baselines, while CCIL still improves over BC.
Limitations
- Corrective labels may confine the agent to a subset of expert-visited states, so they may not yield optimal behavior.
- Relies on the assumption of local Lipschitz continuity in the dynamics; discontinuities are only partly handled.
- Evaluated only in simulation; pixel-based inputs and real robots are left to future work.
Experimental setup
Datasets: D4RL MetaWorld expert demonstrations F1tenth simulator demonstrations gym-pybullet-drones demonstrations (PID-generated)
Benchmarks: Pendulum Discontinuous Pendulum F1tenth racing gym-pybullet-drones hover/circle/fly-through MuJoCo Hopper/Walker2D/Ant/HalfCheetah MetaWorld CoffeePull/ButtonPress/CoffeePush/DrawerClose
Metrics: average return/score success rate L2 label error vs theoretical bound
Baselines: BC NoiseBC MILO MOREL
Builds on: behavior cloning NoiseBC MILO MOREL spectral normalization
Closest papers by meaning
- Action Chunking and Data Augmentation Yield Exponential Improvements in Behavior Cloning for Continuous Spaces (ICLR 2026 · similarity 0.91)
- Imitation Learning by Estimating Expertise of Demonstrators (ICML 2022 · similarity 0.89)
- Markov Balance Satisfaction Improves Performance in Strictly Batch Offline Imitation Learning (AAAI 2025 · similarity 0.88)
- Efficient Imitation under Misspecification (ICLR 2025 · similarity 0.88)
- GO-DICE: Goal-Conditioned Option-Aware Offline Imitation Learning via Stationary Distribution Correction Estimation (AAAI 2024 · similarity 0.88)
- Robust Visual Imitation Learning with Inverse Dynamics Representations (AAAI 2024 · similarity 0.88)
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.