Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
$\textit{HiMaCon:}$ Discovering Hierarchical Manipulation Concepts from Unlabeled Multi-Modal Data
Authors: Ruizhe Liu, Pei Zhou, Qian Luo, Li Sun, Jun CEN, Yibing Song, Yanchao Yang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical evaluation across simulated benchmarks and real-world deployments demonstrates significant performance improvements with our concept-enhanced policies. Our experiments across both simulated benchmark tasks and real-world robot deployments demonstrate that policies enhanced with these manipulation concepts consistently outperform conventional approaches, particularly in challenging scenarios requiring adaptation to novel objects, unexpected obstacles, and environmental variations (Fig. 1). |
| Researcher Affiliation | Collaboration | 1HKU Musketeers Foundation Institute of Data Science, The University of Hong Kong 2Department of Electrical and Electronic Engineering, The University of Hong Kong 3DAMO Academy, Alibaba Group 4Transcengram EMAIL EMAIL, EMAIL |
| Pseudocode | Yes | B Pseudocode Here we provide pseudocode for (i) Deriving subprocess from manipulation concept latents (Alg. 1). (ii) Manipulation concept disocovery training process of our method (Alg. 2). |
| Open Source Code | Yes | Code is available at: https://github.com/zrllrz/Hi Ma Con. |
| Open Datasets | Yes | We conduct experiments using the LIBERO benchmark [30], a comprehensive platform for robotic learning built on Robosuite [75]. We deploy concept-enhanced policies on a Mobile ALOHA robot [16]. We utilized the Bridge Data V2 dataset [60]. |
| Dataset Splits | Yes | LIBERO-90: A diverse collection of 90 manipulation tasks serving as our primary training domain for concept discovery and initial policy learning. Each task includes a natural language description and 50 expert demonstrations. LIBERO-LONG: 10 novel long-horizon tasks, each composed of two LIBERO-90 tasks in sequence, designed to evaluate transfer to more complex task structures. LIBERO-GOAL: 10 tasks in an entirely novel environment unseen during concept discovery, used to evaluate the generalization of learned concepts to unfamiliar contexts. We conduct experiments on LIBERO-90 tasks using the diffusion policy. As shown in Tab. 9, incorporating manipulation concepts consistently improves policy performance compared to settings without them, even under restricted data conditions. This demonstrates that learning and leveraging manipulation concepts can make imitation learning more data-efficient and effective. Table 9: Performance under data constraints. Success rates of diffusion policies with and without manipulation concept enhancement, evaluated on LIBERO-90 (L90-90). In each setting, the number of demonstrations per task available for training both the manipulation concept encoder and the policy is limited as indicated. 50 demos/task 25 demos/task 10 demos/task |
| Hardware Specification | Yes | This training setup is compatible with GPUs such as the Ge Force RTX 3090 or 4090. However, we leverage the A800 GPU for improved efficiency, completing the training process in 1.5 days. To study generalization capabilities, we deploy concept-enhanced policies on a Mobile ALOHA robot [16] in cleaning cup tasks (Fig. 5). |
| Software Dependencies | No | The paper refers to various software components and models (e.g., ACT [71], Diffusion Policy [9], VAE encoder from stable diffusion [49], Adam W optimizer, CLIP [46], DINOv2 [43], Decision NCE-T model from github.com/2toinf/ Decision NCE, OpenAI CLIP model from github.com/openai/ CLIP, dinov2-small DINOv2 model from huggingface.co/facebook/ dinov2-small) but does not provide explicit version numbers for general programming languages or libraries (e.g., Python, PyTorch/TensorFlow) which are key software components for reproducibility. |
| Experiment Setup | Yes | We train the manipulation concept discovery process for 200,000 iterations with a batch size of 512. Each item in the batch is a segment of demonstration with a fixed length of Tcontext = 60. The training process uses the Adam W optimizer with a weight decay of 0.001 and momentum parameters β1 = 0.9 and β2 = 0.95. The base learning rate is set to 0.001. Initially, the model is trained with a 100-iteration warmup phase, during which the learning rate increases linearly from 0.0001 to 0.001. After the warmup, the model is trained for the remaining iterations using a cosine decay schedule, gradually reducing the learning rate back to 0.0001. |