Research map / Reinforcement learning / Imitation learning

DynaMo: In-Domain Dynamics Pretraining for Visuo-Motor Control

NeurIPS 2024 · Reinforcement learning › Imitation learning · approach: Behavior cloning theory and models · 2 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 DynaMo, a self-supervised in-domain visual pretraining method that jointly learns an encoder with latent inverse and forward dynamics models on demonstration image sequences, improving downstream imitation policies on simulated and real manipulation tasks.

Problem

Visuo-motor imitation learning needs many expert demonstrations because visual representations are either pretrained out-of-domain or learned only via behavior cloning; existing in-domain self-supervised objectives ignore action-based temporal causality.

Approach

Jointly train an image encoder with a latent inverse dynamics model (inferring unobserved transition latents from consecutive frame embeddings) and a forward dynamics model predicting the next frame embedding, with stop-gradient (SimSiam) or EMA targets and covariance regularization, requiring no augmentations, contrastive sampling, or ground-truth actions.

Key results

• DynaMo reaches 0.65 on Block Pushing and 0.66 on Push-T versus best prior self-supervised baselines of 0.52 (RPT) and 0.57 (MoCo-v3).
• On real Allegro hand, DynaMo gets 7/10, 5/10, 9/10 on sponge, tea, microwave, outperforming the best baseline by 43% on average.
• Across policy classes on Push-T, DynaMo achieves 0.66 with VQ-BeT, 0.73 with Diffusion Policy, 0.35 with MLP, and 0.12 with kNN, the best self-supervised result in each column.

Measured effects reported in the paper

Surprising findings

• Using ground-truth action labels during visual encoder training does not improve downstream policy performance, which the authors attribute to the policy already having access to the actions.
• Removing the stop-gradient or momentum target causes representation collapse, with near-zero performance on Franka Kitchen and LIBERO.

Limitations

• Evaluated only on lab settings; authors note more research is needed for manipulation outside labs.
• Authors did not test training on larger unlabeled datasets; they suggest this could improve generalization.
• Some gains are task-dependent; e.g., on Franka Kitchen DynaMo matches rather than beats the best baseline (BYOL 3.75 vs DynaMo 3.64).

Experimental setup

Datasets: Franka Kitchen Block Pushing Push-T LIBERO Goal Allegro Manipulation (self-collected) xArm Kitchen (self-collected)

Benchmarks: Franka Kitchen Block Pushing Push-T LIBERO Goal Allegro Manipulation xArm Kitchen

Metrics: mean completed tasks mean blocks in targets final target coverage average success rate success count out of trials

Baselines: ImageNet R3M VC-1 MVP BYOL BYOL-T MoCo-v3 RPT TCN MAE

Builds on: SimSiam MoCo-v3 VQ-BeT R3M nanoGPT

Closest papers by meaning

• Latent Policy Barrier: Learning Robust Visuomotor Policies by Staying In-Distribution (NeurIPS 2025 · similarity 0.89)
• Latent Diffusion Planning for Imitation Learning (ICML 2025 · similarity 0.89)
• Robots Pre-train Robots: Manipulation-Centric Robotic Representation from Large-Scale Robot Datasets (ICLR 2025 · similarity 0.89)
• VidMan: Exploiting Implicit Dynamics from Video Diffusion Model for Effective Robot Manipulation (NeurIPS 2024 · similarity 0.88)
• Predictive Inverse Dynamics Models are Scalable Learners for Robotic Manipulation (ICLR 2025 · similarity 0.88)
• FreqPolicy: Frequency Autoregressive Visuomotor Policy with Continuous Tokens (NeurIPS 2025 · similarity 0.88)