GORDON: Graph-based Object-centric Rewards for Decomposition of Long-Horizon Manipulation
AI Summary
GORDON learns dense reinforcement-learning rewards from action-free video demonstrations by representing scenes as graphs of detected objects and spatial relations. A self-supervised graph neural network maps these graphs into a task-aligned latent space, while activity-aware pooling emphasizes relevant objects and masks robot-dominated motion. Distances to demonstrated goal configurations define progress rewards. Temporal reward profiles expose stage-wise object transitions, enabling automatic subtask discovery and sequential policy composition. On seven tasks from MAGICAL and ManiSkill3, the authors report a 74.4% average success rate on long-horizon tasks, about 35 percentage points above the best learned baseline and 25 points above an oracle-based comparison.
Why it's worth reading
Long-horizon robot learning is moving beyond manually labeled subtasks; this paper offers an object-centric video-reward pipeline with reported benchmark gains and automatic decomposition.
Deep Read
What happened
Original facts: The paper introduces GORDON, a method for learning dense rewards from action-free video demonstrations and using reward trajectories to discover stages and subtasks in long-horizon manipulation. Experiments are reported on seven manipulation tasks from MAGICAL and ManiSkill3, with a 74.4% average success rate on the long-horizon tasks.
Core technology
Original facts: Each visual scene is represented as a graph containing detected objects and spatial relations. A graph neural network is trained self-supervised to embed these graphs into a task-aligned latent space. Activity-aware weighted pooling emphasizes task-relevant objects while masking robot-dominated motion. Dense reward is computed from latent-space distances between the current state and demonstrated goal configurations. Temporal reward profiles are then used for stage discovery, followed by subtask-specific reward and policy training and sequential policy composition.
Key evidence & numbers
Original facts: The evaluation covers seven tasks across MAGICAL and ManiSkill3. The authors report a 74.4% average success rate for long-horizon tasks, approximately 35 percentage points above the best learned baseline and approximately 25 points above an oracle comparison. The supplied abstract does not specify per-task variance, the number of runs, confidence intervals, or the complete definitions of the baselines.
Why it matters
Analysis: Long-horizon reinforcement learning is constrained by sparse rewards, visual variation, and the difficulty of identifying subtask boundaries. GORDON combines object-centric representation, reward learning, and task decomposition in one pipeline. Its distinctive claim is that reward dynamics can provide both a learning signal and a source of stage structure. If replicated across broader visual and robotic settings, this could reduce the cost of manually annotating subtasks.
Practical impact
Analysis: Researchers could use the framework to derive object-level rewards from demonstrations and train specialized policies for automatically discovered segments. In an engineering system, object detection, relation estimation, and activity weighting would likely be central components. The approach appears most naturally suited to tasks with identifiable object interactions, such as rearrangement, assembly, and multi-step transport. The paper does not establish that it replaces real-robot data collection, controller design, or safety validation.
Limitations & uncertainty
Original facts: The abstract does not describe the object detector, graph-construction failure modes, demonstration volume, training cost, viewpoint generalization, or real-robot validation. Analysis: Object-centric representations may degrade under occlusion, transparent objects, dense contact, or changes in object categories. Latent-space distance may also fail to correspond monotonically to task progress. Unverified inference: The reported gains may depend on task selection, baseline implementation, the oracle definition, or evaluation protocol; reproducibility requires checking the full paper and any released code.
Original sources
- arXiv abstract page
- Source: arXiv
- Published: 2026-08-04T14:43:41.000Z