Research map / Reinforcement learning / Visual and representation RL
$\textit{Hyper-GoalNet}$: Goal-Conditioned Manipulation Policy Learning with HyperNetworks
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 Hyper-GoalNet, which uses an optimization-inspired hypernetwork to generate goal-specific visuomotor policy weights from a single goal image, with latent dynamics and distance constraints shaping the encoder, evaluated in Robosuite/MimicGen and on a real 7-DoF arm.
Problem
Goal-conditioned visuomotor manipulation policies with fixed-parameter networks that concatenate goal and current observations generalize poorly, especially under high object-pose variability and contact-rich tasks.
Approach
Treat the goal image as the generator of the policy's own parameters via an optimization-inspired hypernetwork (iterative refinement of weights), and shape the latent space with a forward dynamics loss and a monotonic goal-distance hinge loss so that the hypernetwork receives clear progress signals.
Key results
- On Robosuite Coffee (avg over d0-d2) Hyper-GoalNet reaches 0.77 vs 0.55 for C-BeT, the strongest baseline in the main table.
- On long-horizon Kitchen, Hyper-GoalNet achieves 0.90 average vs 0.74 for C-BeT.
- On real-robot tasks, Hyper-GoalNet succeeds 14/15, 15/15, 14/15, 15/15 across pick&place, pull, stack, sweep, while C-BeT achieves at most 8/15.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use goal-conditioned hypernetwork policy generation instead of fixed-parameter C-BeT | goal-conditioned manipulation policies · Coffee Avg. | success rate: 0.55 -> 0.77 | Helped |
| use goal-conditioned hypernetwork policy generation instead of fixed-parameter C-BeT | goal-conditioned manipulation policies · Mug cleanup Avg. | success rate: 0.40 -> 0.62 | Helped |
| use goal-conditioned hypernetwork policy generation instead of fixed-parameter C-BeT | goal-conditioned manipulation policies · Three-piece assembly Avg. | success rate: 0.01 -> 0.25 | Helped |
| use goal-conditioned hypernetwork policy generation instead of fixed-parameter C-BeT | goal-conditioned manipulation policies · Threading Avg. | success rate: 0.32 -> 0.46 | Helped |
| use goal-conditioned hypernetwork policy generation instead of fixed-parameter C-BeT | goal-conditioned manipulation policies · Coffee preparation Avg. | success rate: 0.43 -> 0.65 | Helped |
| use goal-conditioned hypernetwork policy generation instead of fixed-parameter C-BeT | goal-conditioned manipulation policies · Kitchen Avg. | success rate: 0.74 -> 0.90 | Helped |
| remove latent space shaping (ablation) | Hyper-GoalNet · Coffee Avg. | success rate: 0.77 -> 0.31 | Hurt |
| unfreeze the visual encoder at epoch 0 instead of after 20 epochs (ablation) | Hyper-GoalNet · Coffee Avg. | success rate: 0.77 -> 0.51 | Hurt |
| use distance to the start image instead of goal-relative distance (ablation) | Hyper-GoalNet · Coffee Avg. | success rate: 0.77 -> 0.45 | Hurt |
| replace Euclidean distance with cosine distance (ablation) | Hyper-GoalNet · Coffee Avg. | success rate: 0.77 -> 0.59 | Hurt |
| replace the optimization-inspired hypernetwork with HyperZero using ScalarInit (ablation) | Hyper-GoalNet · Coffee task Avg. | success rate: 77 -> 16 | Hurt |
| replace the optimization-inspired hypernetwork with HyperZero using Bias-Init (ablation) | Hyper-GoalNet · Coffee task Avg. | success rate: 77 -> 16 | Hurt |
Surprising findings
- GCBC and Play-LMP achieve 0% success on nearly all simulated tasks, which the authors attribute to overfitting of likelihood-based objectives.
- Applying the latent shaping to C-BeT improves it (0.69 avg on Coffee ablation), but it still trails the full method.
Limitations
- Relies on a well-structured latent space that requires demonstrations with clear goal progression.
- Out-of-distribution goals can cause the hypernetwork to generate erratic policies.
- Offline-trained method without explicit safety guarantees.
Experimental setup
Datasets: MimicGen
Benchmarks: Robosuite Coffee Mug-cleanup Three piece assembly Threading Nut assembly Coffee preparation Kitchen real-robot sweep/pick&place/pull/stack
Metrics: success rate over 50 rollouts accuracy and recall for goal completion detection
Baselines: GCBC Play-LMP C-BeT MimicPlay HyperZero BeT
Builds on: HyperZero R3M C-BeT MimicGen
Closest papers by meaning
- GOALNET: Interleaving Neural Goal Predicate Inference with Classical Planning for Generalization in Robot Instruction Following (AAAI 2024 · similarity 0.89)
- Efficient Robotic Policy Learning via Latent Space Backward Planning (ICML 2025 · similarity 0.87)
- Goal-Oriented Skill Abstraction for Offline Multi-Task Reinforcement Learning (ICML 2025 · similarity 0.86)
- Multi-Stage Manipulation with Demonstration-Augmented Reward, Policy, and World Model Learning (ICML 2025 · similarity 0.86)
- Hierarchical Equivariant Policy via Frame Transfer (ICML 2025 · similarity 0.86)
- Stabilizing Contrastive RL: Techniques for Robotic Goal Reaching from Offline Data (ICLR 2024 · 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.