Research map / Vision-language and multimodal / Robot action and navigation
ThinkAct: Vision-Language-Action Reasoning via Reinforced Visual Latent Planning
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.
ThinkAct trains a Qwen2.5-VL-7B reasoning MLLM with GRPO using action-aligned visual rewards (goal completion and DTW trajectory consistency) to produce visual plan latents that condition a DiT-based action policy.
Problem
Vision-language-action models map observations directly to actions without explicit multi-step reasoning, limiting long-horizon planning and adaptation; QA-style RL rewards do not ground reasoning in executable action.
Approach
Reinforce an MLLM's embodied reasoning with rewards computed from predicted 2D gripper trajectories (start/end goal reward and DTW trajectory reward), then compress the reasoning into a visual plan latent that conditions a separate diffusion-transformer action model via a Q-Former projector.
Key results
- ThinkAct achieves 71.5%, 65.1%, and 43.8% overall on SimplerEnv Google-VM, Google-VA, and Bridge-VM, and 84.4% overall on LIBERO, outperforming DiT-Policy and CoT-VLA (83.9%).
- On embodied reasoning, ThinkAct reaches 48.2% on EgoPlan-Bench2 and 59.8 BLEU overall on RoboVQA, beating the second-best method by 2.5% and 4.1 BLEU.
- Few-shot adaptation with 10 demonstrations per task surpasses Magma by 7.3% on LIBERO-Goal and 9.5% on LIBERO-Spatial.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| condition DiT-Policy action execution on reasoning-guided visual plan latents | ThinkAct · SimplerEnv Google Visual Matching, Google Variant Aggregation, and Bridge Visual Matching | success rate: 15.5%, 16.9%, and 11.4% | Helped |
| use ThinkAct rather than CoT-VLA | Robot manipulation models · LIBERO Overall | success rate: 84.4% vs. 83.9% | Helped |
| use ThinkAct rather than the strongest reported baseline | Embodied reasoning models · EgoPlan-Bench2 Overall | accuracy: 2.5% | Helped |
| use ThinkAct rather than the strongest reported baseline | Embodied reasoning models · RoboVQA Overall | BLEU: 4.1 BLEU score | Helped |
| add the trajectory reward to reinforced fine-tuning (ablation) | ThinkAct reasoning model · Embodied reasoning and planning benchmarks | performance | Helped |
| add the goal reward to reinforced fine-tuning (ablation) | ThinkAct reasoning model · Embodied reasoning and planning benchmarks | performance | Helped |
| add action-aligned visual goal and trajectory rewards instead of relying only on QA-style reward (ablation) | ThinkAct reasoning model · Embodied multi-step planning | performance | Helped |
| apply RL fine-tuning after SFT cold-start (ablation) | ThinkAct reasoning model · Embodied reasoning tasks | reasoning performance | Helped |
| vary actions per reasoning step from N=25 to N=100 (ablation) | ThinkAct action adaptation · LIBERO | average success rate: 84.0%, 84.6%, 84.4%, and 83.7% | Mixed |
| set actions per reasoning step to N=100 rather than N=75 (ablation) | ThinkAct action adaptation · LIBERO | average success rate: 84.4% to 83.7% | Hurt |
| use N=25 instead of N=75 for more frequent reasoning updates (ablation) | ThinkAct action adaptation · LIBERO | average success rate: 84.0% vs. 84.4% | No clear effect |
Surprising findings
- Removing both action-aligned rewards leaves only marginal gains over the SFT baseline on the reasoning benchmarks, indicating QA-style rewards alone are insufficient.
- ThinkAct takes about 17% longer execution time than OpenVLA due to autoregressive reasoning.
Limitations
- Inherits hallucinations from pretrained MLLMs, leading to incorrect object or spatial references in plans.
- Inference is slower than end-to-end OpenVLA due to autoregressive reasoning.
- Evaluation largely in simulation (SimplerEnv, LIBERO) with a single third-person 224x224 image observation.
Experimental setup
Datasets: Open X-Embodiment Something-Something V2 RoboVQA EgoPlan-IT EgoPlan-Val Video-R1-CoT LLaVA-Video-178K RoboFail LIBERO
Benchmarks: SimplerEnv LIBERO EgoPlan-Bench2 RoboVQA OpenEQA
Metrics: task success rate accuracy BLEU LLM-based score
Baselines: Octo-Base RT1-X OpenVLA DiT-Policy TraceVLA CoT-VLA Magma GPT-4V LLaVA-Video InternVL2.5 InternVL3 NVILA
Builds on: GRPO DeepSeek-R1 Diffusion Policy HAMSTER Video-R1 OpenVLA
Closest papers by meaning
- Fast-ThinkAct: Efficient Vision-Language-Action Reasoning via Verbalizable Latent Planning (CVPR 2026 · similarity 0.94)
- SemanticVLA: Towards Semantic Reasoning over Action Memorization via Synergistic Explicit Trace and Latent Action Planning (CVPR 2026 · similarity 0.91)
- Latent Reasoning VLA: Latent Thinking and Prediction for Vision-Language-Action Models (ICML 2026 · similarity 0.91)
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
- EVLP: Learning Unified Embodied Vision-Language Planner with Reinforced Supervised Fine-Tuning (ICLR 2026 · similarity 0.89)
- APIVOT: Adaptive Planning with Interleaved Vision-Language Thoughts (NeurIPS 2026 · similarity 0.89)
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.