Research map / Vision-language and multimodal / Robot action and navigation
Temporal Representation Alignment: Successor Features Enable Emergent Compositionality in Robot Instruction Following
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 Temporal Representation Alignment (TRA), which adds a time-contrastive successor-feature alignment loss to goal- and language-conditioned behavior cloning, yielding zero-shot compositional generalization on robot manipulation and OGBench without planning or RL.
Problem
Learning task representations that let an agent compose previously learned short-horizon goal-reaching and instruction-following skills into unseen multi-step tasks, without explicit planning, hierarchy, or RL.
Approach
Add an InfoNCE temporal alignment loss that aligns current-state representations with representations of future states (geometric discount sampling), shared with goal and language embeddings, so that a plain behavior cloning policy conditioned on these representations can stitch expert behaviors together.
Key results
- On real-world instruction-following compositional tasks, TRA outperforms the best baseline by large margins (e.g., 88.9% on Set B vs 30% for LCBC; 60% on Set D vs 33.3% for Octo).
- TRA beats AWR by about 45% on instruction-following tasks and about 25% on goal-reaching tasks in the real-world evaluation.
- On OGBench, TRA achieves 60.7% on antmaze medium stitch vs 45.5% for GCBC and 52.7% for CRL, and 46.1% on humanoidmaze medium stitch vs 36.2% for CRL.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add temporal representation alignment to behavior cloning | BridgeData language-conditioned policy · task concatenation | success rate: 88.9% vs 30% | Helped |
| add temporal representation alignment to behavior cloning | BridgeData language-conditioned policy · semantic generalization | success rate: 83.3% vs 43.3% | Helped |
| add temporal representation alignment to behavior cloning | BridgeData language-conditioned policy · tasks with dependency | success rate: 60% vs 33.3% | Helped |
| add temporal representation alignment to behavior cloning | BridgeData goal-image-conditioned policy · semantic generalization | success rate: 60% vs 46.7% | Helped |
| add temporal representation alignment to behavior cloning | BridgeData goal-image-conditioned policy · tasks with dependency | success rate: 43.3% vs 20% | Helped |
| add temporal representation alignment rather than use GCBC | OGBench policies · stitch and goal-reaching environments | success rate | Mixed |
| remove temporal alignment loss (ablation) | GCBC-ϕ on OGBench · antmaze medium stitch | success rate: 48.7±(2.7) vs 60.7(±3.0) | Hurt |
| remove temporal alignment loss (ablation) | GCBC-ϕ on OGBench · antsoccer arena stitch | success rate: 1.4±(0.3) vs 17.0(±1.2) | Helped |
| remove temporal alignment loss (ablation) | GCBC-ϕ on OGBench · humanoidmaze medium stitch | success rate: 34.4±(1.7) vs 46.1(±1.9) | Hurt |
| remove temporal alignment loss (ablation) | GCBC-ϕ on OGBench · humanoidmaze large stitch | success rate: 3.5±(1.1) vs 8.6(±1.4) | Hurt |
| remove temporal alignment loss (ablation) | GCBC-ϕ on OGBench · antmaze large navigate | success rate: 16.1±(0.8) vs 35.4(±1.8) | Hurt |
| remove temporal alignment loss (ablation) | GCBC-ϕ on OGBench · cube single noisy | success rate: 8.7±(0.9) vs 9.2(±0.9) | No clear effect |
Surprising findings
- Using the temporal alignment representation alone, without AWR policy weighting, is sufficient; adding AWR does not significantly help and marginally decreases performance.
- Supervised behavior cloning with structured temporal representations can match or exceed offline RL methods on compositional tasks.
Limitations
- Improvements degrade when behavior cloning cannot solve the task at all.
- Gaussian policies struggle with multimodal behavior, causing early grasping or incorrect reaching.
- Not yet validated with transformer or diffusion policies at larger scale.
Experimental setup
Datasets: BridgeDataV2 OGBench
Benchmarks: OGBench BridgeData real-world WidowX evaluation
Metrics: success rate
Baselines: GRIF GCBC LCBC Octo AWR GCIQL GCIVL QRL CRL
Builds on: GRIF CLIP contrastive RL (CRL) advantage-weighted regression Octo
Closest papers by meaning
- RoboDreamer: Learning Compositional World Models for Robot Imagination (ICML 2024 · similarity 0.88)
- Modeling Unseen Environments with Language-guided Composable Causal Components in Reinforcement Learning (ICLR 2025 · similarity 0.88)
- ClevrSkills: Compositional Language And Visual Reasoning in Robotics (NeurIPS 2024 · similarity 0.87)
- Learning with Language-Guided State Abstractions (ICLR 2024 · similarity 0.87)
- $\textit{HiMaCon:}$ Discovering Hierarchical Manipulation Concepts from Unlabeled Multi-Modal Data (NeurIPS 2025 · similarity 0.87)
- Dream to Manipulate: Compositional World Models Empowering Robot Imitation Learning with Imagination (ICLR 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.