Research map / Reinforcement learning / Visual and representation RL

AdaManip: Adaptive Articulated Object Manipulation Environments and Policy Learning

ICLR 2025 · Reinforcement learning › Visual and representation RL · approach: Dexterous manipulation transfer · 1 citations (OpenAlex)

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.

Builds AdaManip, an articulated-object manipulation environment with 9 categories of 277 objects and 5 adaptive mechanisms, and trains a 3D diffusion policy on adaptive demonstrations containing failure-and-recovery trials.

Problem

Articulated objects often have hidden internal states (e.g., lock state, joint limits, rotation direction) that cannot be inferred from vision, so manipulation requires adaptive trial-and-error; existing datasets and environments mostly cover simple mechanisms inferable from appearance.

Approach

Collect demonstrations that include a failed trial followed by recovery (optimal under partial observation) and train a DDPM-based 3D diffusion policy conditioned on observation and action history, so the policy's multimodal distribution collapses adaptively after feedback.

Key results

• Ours achieves 82.53% average success across the nine simulated categories versus 53.57% for DP3 and 51.19% for ACT.
• Ours w/o adaptive (static demonstrations) reaches 54.29% average, showing that adaptive failure-recovery demonstrations drive much of the gain.
• Real-world success: 8/10 open bottle, 7/10 open microwave, 5/10 open safe, 5/10 open pressure cooker.

Measured effects reported in the paper

Surprising findings

• Increasing the number of repeated adaptive trials in demonstrations degrades performance (0 trials 0.8714 vs 1 trial 0.9364 and 6 trials 0.4428 for Bottle), since extra trials add distributional complexity.
• DP3 without adaptive demonstrations performs close to ACT, and the sampling-based planner performs worst, below affordance-based VAT-Mart on several categories.

Limitations

• Dataset covers only 9 categories and 5 mechanism types, with no deformable objects.
• Real-world success rates for safe and pressure cooker are only 5/10, and real-world evaluation uses few trials per object.
• Repeated adaptive trials in demonstrations hurt performance, so the one-trial design is a hand-tuned restriction rather than a general solution.

Experimental setup

Datasets: AdaManip adaptive articulated object dataset (9 categories, 277 objects)

Benchmarks: AdaManip simulation environment (9 categories) real-world Franka tasks (open bottle, microwave, safe, pressure cooker)

Metrics: success rate

Baselines: VAT-Mart AdaAfford Sampling ACT DP3 Ours w/o adaptive

Builds on: Diffusion Policy DP3 IsaacGym PointNet++ DDPM

Closest papers by meaning

• General Articulated Objects Manipulation in Real Images via Part-Aware Diffusion Process (NeurIPS 2024 · similarity 0.88)
• JODA: Composable Joint Dynamics for Articulated Objects (NeurIPS 2026 · similarity 0.87)
• GMAP: Generalized Manipulation of Articulated Objects in Robotic Using Pre-trained Model (AAAI 2025 · similarity 0.87)
• Data Scaling Laws in Imitation Learning for Robotic Manipulation (ICLR 2025 · similarity 0.86)
• ManipTrans: Efficient Dexterous Bimanual Manipulation Transfer via Residual Learning (CVPR 2025 · similarity 0.86)
• Omnigrasp: Grasping Diverse Objects with Simulated Humanoids (NeurIPS 2024 · similarity 0.86)