Research map / Reinforcement learning / Visual and representation RL
SOLD: Slot Object-Centric Latent Dynamics Models for Relational Manipulation Learning from Pixels
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 SOLD, an object-centric model-based RL method that learns slot-based latent dynamics from pixels (SAVi + OCVP-style transformer dynamics + Slot Aggregation Transformer for reward/value/actor), outperforming DreamerV3 and TD-MPC2 on relational robotic manipulation tasks.
Problem
Model-based RL from pixels typically relies on holistic state representations, which hinders relational reasoning and manipulation over multiple objects; object-centric world models for pixel-based RL are largely unexplored.
Approach
Combine a pretrained SAVi slot encoder-decoder with an action-conditioned, autoregressive object-centric dynamics transformer (OCVP-style sequential attention), and a Slot Aggregation Transformer with output and register tokens (ALiBi positions) that serves as backbone for reward, critic and actor, trained with Dreamer-style imagination.
Key results
- On Distinct tasks, SOLD averages 74.6% success vs 33.1% for DreamerV3 and 17.3% for TD-MPC2.
- On Specific tasks, SOLD averages 89.4% success vs 81.7% (DreamerV3) and 84.2% (TD-MPC2).
- SOLD reaches 100% success on Meta-World Button-Press and Hammer, and returns of 497 (Cartpole-Balance) and 645 (Finger-Spin).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SOLD instead of TD-MPC2 | visual robotic control · Specific tasks average | success rate (%): 84.2 ± 9 -> 89.4 ± 6 | Helped |
| Use SOLD instead of DreamerV3 | visual robotic control · Distinct tasks average | success rate (%): 33.1 ± 23 -> 74.6 ± 13 | Helped |
| Replace object-centric SAVi encoder-decoder with a convolutional autoencoder (ablation) | SOLD · Reach-Specific | success rate (%): 97.9 ± 0 -> 83.2 ± 2 | Hurt |
| Replace object-centric SAVi encoder-decoder with a convolutional autoencoder (ablation) | SOLD · Reach-Rel. | success rate (%): 91.1 ± 2 -> 39.2 ± 3 | Hurt |
| Replace object-centric SAVi encoder-decoder with a convolutional autoencoder (ablation) | SOLD · Push-Specific | success rate (%): 82.8 ± 2 -> 75.2 ± 2 | Hurt |
| Replace object-centric SAVi encoder-decoder with a convolutional autoencoder (ablation) | SOLD · Pick-Specific | success rate (%): 85.8 ± 7 -> 22.9 ± 11 | Hurt |
| Replace object-centric SAVi encoder-decoder with a convolutional autoencoder (ablation) | SOLD · Reach-Distinct | success rate (%): 91.8 ± 1 -> 11.3 ± 1 | Hurt |
| Replace object-centric SAVi encoder-decoder with a convolutional autoencoder (ablation) | SOLD · Reach-Gr. | success rate (%): 69.6 ± 2 -> 5.1 ± 1 | Hurt |
| Replace object-centric SAVi encoder-decoder with a convolutional autoencoder (ablation) | SOLD · Push-Distinct | success rate (%): 80.6 ± 5 -> 10.5 ± 1 | Hurt |
| Replace object-centric SAVi encoder-decoder with a convolutional autoencoder (ablation) | SOLD · Pick-Distinct | success rate (%): 56.4 ± 25 -> 0.7 ± 0 | Hurt |
| Fine-tune SAVi encoder-decoder during training rather than keeping it frozen (ablation) | SOLD · Pick tasks | reconstruction quality | Helped |
Surprising findings
- The non-object-centric CNN baseline, scaled to about 60M parameters, performs poorly on Distinct tasks (6.9% average), showing object structure matters more than capacity.
- DreamerV3 outperforms SOLD on the Push-Specific task (97.1% vs 82.8%).
Limitations
- The world model is deterministic, which may hurt in stochastic environments.
- SAVi may not scale to complex real-world visual data.
- Evaluation is limited to simulated tasks with a fixed set of object-centric environments.
Experimental setup
Datasets: MuJoCo-based object-centric robotic environments (Reach, Push, Pick; Specific and Distinct variants) Meta-World Button-Press and Hammer DM-Control Cartpole-Balance and Finger-Spin
Benchmarks: Reach-Specific Reach-Relative Push-Specific Pick-Specific Reach-Distinct Reach-Groups Push-Distinct Pick-Distinct Meta-World Button-Press Meta-World Hammer DM-Control Cartpole-Balance DM-Control Finger-Spin
Metrics: success rate (%) episode return
Baselines: DreamerV3 TD-MPC2 Non-Object-Centric CNN baseline (w/o OCE)
Builds on: SAVi OCVP DreamerV3 TD-MPC2 ALiBi
Closest papers by meaning
- Entity-Centric Reinforcement Learning for Object Manipulation from Pixels (ICLR 2024 · similarity 0.88)
- LaST-R1: Reinforcing Robotic Manipulation via Adaptive Physical Latent Reasoning (NeurIPS 2026 · similarity 0.88)
- PlaySlot: Learning Inverse Latent Dynamics for Controllable Object-Centric Video Prediction and Planning (ICML 2025 · similarity 0.87)
- Learning Interactive World Model for Object-Centric Reinforcement Learning (NeurIPS 2025 · similarity 0.87)
- EC-Diffuser: Multi-Object Manipulation via Entity-Centric Behavior Generation (ICLR 2025 · similarity 0.87)
- AffordDP: Generalizable Diffusion Policy with Transferable Affordance (CVPR 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.