Source-linked AI summary
Learning to Explore Hidden Kinematics for Articulated Object Manipulation
Ruiyao Liu, Boshu Lei, Zhuoyang Pan, Kostas Daniilidis
TL;DR
Articulated-object kinematics can remain ambiguous from vision, while extending greedy action selection over time requires unknown contact and inertial dynamics. The paper learns an interaction policy conditioned on a posterior-derived articulation flow and rewards entropy reduction; it outperforms existing approaches across door and drawer tasks and reports 61.7% success on ArticuRiddle.
Problem
Visual appearance can leave articulated-object kinematics ambiguous, and greedy action selection cannot be extended over a horizon without unknown contact and inertial dynamics.
Source: §I · p. 1Method
The method updates a belief over kinematics from observed part motion, renders it as a pointwise articulation flow for the policy, and rewards posterior entropy reduction during reinforcement learning.
Source: §I · pp. 1–2Results
The method outperforms existing approaches across door and drawer tasks and achieves 61.7% success on ArticuRiddle.
Source: §I · p. 2Takeaways & Limitations
Interaction lets the articulation estimate and its corresponding flow field be progressively refined on real-world articulated objects through zero-shot transfer.
Source: §I · p. 2
Abstract
from arXiv · showhide
The kinematics of an articulated object is often ambiguous from vision alone. Interaction resolves the ambiguity, and active perception methods exploit this by searching for the single action that most sharpens a belief over the kinematic parameters at each step. Such greedy search cannot be extended over a horizon without forward models of the contact and inertial dynamics, which are themselves unknown. We instead amortize action selection into training. We maintain a belief distribution over joint type and parameters, initialized from a generative prior and updated by Bayesian filtering on the observed part motion. To condition the policy on this belief, we render it as a per-point articulation flow field, the motion that the current posterior predicts for every point on the object. Carrying the inductive bias of articulated motion, this representation generalizes better than a latent encoding of the belief or flow tracked from observation. We train the policy with reinforcement learning, rewarding the entropy that each interaction removes from the posterior, so that informative exploration becomes learned behavior rather than a search at every step. Our method outperforms previous approaches across door and drawer manipulation on the PartManip benchmark, and reaches 61.7% success on ArticuRiddle, a new dataset of objects whose appearance implies the wrong articulation, against 44.4% for the best previous method. Project Website: https://hiddenkinematics.github.io/
I. INTRODUCTION
Visual appearance can leave an articulated object’s joint ambiguous, while interaction reveals it through the resulting part motion. The paper learns a policy that uses an evolving articulation belief to seek informative motion while manipulating.
I. INTRODUCTION
- Articulated-object geometry can leave different joint structures visually indistinguishable, so resolving the ambiguity requires evidence from interaction.The introduction frames the problem as choosing actions under a belief over kinematic hypotheses so the resulting motion is informative.Source: §I · p. 1
- Prior active-perception methods search for an informative next action, but extending that search over a horizon requires unknown contact and inertial dynamics.The paper instead trains information-seeking behavior rather than scoring every action at execution time.Source: §I · p. 1
- The framework maintains and updates a kinematic belief from observed part motion, then renders it as a pointwise articulation flow that conditions the policy.The flow represents motion predicted by the current belief, while entropy reduction provides a reward for informative interaction alongside task reward.Source: §I · pp. 1–2
- The method outperforms existing approaches across door and drawer tasks and achieves 61.7% success on ArticuRiddle, compared with 32.4% for PartManip.The paper also reports zero-shot transfer to real-world articulated objects, where interaction progressively refines the articulation estimate and flow field.Source: §I · p. 2
- The paper introduces ArticuRiddle for visually ambiguous articulation and reports improvements in simulation plus zero-shot transfer to real-world objects.ArticuRiddle evaluates manipulation when visual appearance does not reliably reveal the articulation.Source: §I · p. 2
II. RELATED WORK
Prior work estimates articulation or selects informative actions, but existing active-perception methods plan only the next action. This method initializes articulation hypotheses from a generative prior and incorporates belief information into manipulation.
II. RELATED WORK
- Prior articulation estimates can leave policies vulnerable when visually similar objects have different kinematics, while posterior-derived flow is recomputed after interaction.Prior work conditions policies on affordance maps, motion fields, or latents; this approach derives flow from the current posterior.Source: §II · p. 2
- Active-perception methods use posterior entropy or heuristics to select informative actions, but their planning is myopic.Reinforcement learning instead optimizes expected information return over a trajectory, amortizing action selection into training.Source: §II · p. 2
III. METHOD
- The task estimates an unknown articulation online by acting on a target part and using its resulting motion as evidence.The method pipeline is summarized in Fig. 2.Source: §III · p. 2
1) Generating Hypothetical Articulations:
- NAP samples candidate joint types and parameters from the initial observation to form particles representing uncertainty over articulation.The candidates include revolute, prismatic, and fixed joints.Source: 1) Generating Hypothetical Articulations · p. 2
- Each particle’s predicted per-point flow is weighted and concatenated with the current point cloud to condition the policy.After acting, observed motion updates particle weights through a Bayes filter; entropy reduction is combined with task reward for PPO training.Source: Figure 2 · p. 3, 1) Generating Hypothetical Articulations · p. 3
2) Updating Hypotheses from Interaction:
After each action, the method compares observed target-part motion with predictions from candidate articulation hypotheses. Better-fitting hypotheses receive greater likelihood, and the posterior is updated recursively.
2) Updating Hypotheses from Interaction:
- Observed motion is represented by tracked target-part points before and after the action, then evaluated against each articulation hypothesis.Revolute hypotheses predict rotation, prismatic hypotheses predict translation, and fixed hypotheses predict negligible motion.Source: 2) Updating Hypotheses from Interaction · p. 3
- The method estimates joint displacement from the same point correspondences and scores each hypothesis by how well its kinematic constraint explains them.The likelihood assumes isotropic Gaussian observation noise; better-fitting hypotheses receive larger likelihood weights.Source: 2) Updating Hypotheses from Interaction · p. 3
- Bayesian updating combines the motion likelihood with the belief accumulated from earlier interactions, then normalizes posterior weights for resampling.Particle filtering implements this recursive inference with importance weighting and resampling.Source: 2) Updating Hypotheses from Interaction · p. 3
B. Flow Fields from Belief Distribution
The policy receives a per-point motion field derived from the current articulation posterior. Recomputing that field after each interaction carries accumulated evidence into the observation.
B. Flow Fields from Belief Distribution
- The current posterior is converted into a per-point motion field using the kinematic constraints represented by its particles.Revolute motion follows a circle tangent, while prismatic motion points along the joint axis.Source: B. Flow Fields from Belief Distribution · p. 4
- Joint-type fields are normalized by their mean magnitude and combined pointwise according to posterior belief, then appended to point features.The augmented point feature changes from R6 to R9.Source: B. Flow Fields from Belief Distribution · p. 4
- Recomputing the field from the posterior at each step gives the policy an implicit memory of prior interactions without requiring it to retain interaction history.The representation addresses the sensitivity of direct planning to estimation errors and the burden of inferring articulation from history alone.Source: B. Flow Fields from Belief Distribution · pp. 3–4
C. Reinforcement Learning for Informative Interaction
The method uses the evolving articulation posterior to reward informative interaction, combining uncertainty reduction with task reward so the policy learns to explore while manipulating.
C. Reinforcement Learning for Informative Interaction
- Belief entropy keeps joint-type uncertainty exact while estimating parameter uncertainty separately within each type branch.Prismatic particles carry an axis direction, revolute particles carry an axis and pivot, and fixed particles contribute only type uncertainty.Source: C. Reinforcement Learning for… · p. 4
- A kernel-density estimate of particle positions measures branch entropy because particle-weight entropy reflects surviving hypothesis count, not their spatial spread.The branch estimate uses a Gaussian kernel on each particle and a Scott-rule bandwidth.Source: C. Reinforcement Learning for… · p. 4
- The information reward is the posterior entropy drop caused by each observation, and it vanishes once the belief stops changing.This avoids rewarding persistent uncertainty or surprise after the posterior has converged.Source: C. Reinforcement Learning for… · p. 4
- PPO combines the information reward with part-aware task reward, while a distilled vision policy is further trained with reinforcement learning and the expert as a regularizer.The task reward includes gripper alignment, distance to the handle centre, and a one-time success reward when the part is sufficiently open.Source: C. Reinforcement Learning for… · p. 4
A. Simulation Settings
The evaluation uses simulated door and drawer manipulation benchmarks, including objects designed to preserve appearance while altering geometric or kinematic cues, and compares the method with prior planning and learned-policy approaches.
A. Simulation Settings
- Simulation uses an IsaacGym Franka Panda arm with a parallel gripper, a fused 4,000-point colored cloud from three depth cameras, and a 41-dimensional proprioceptive vector.The policy predicts an end-effector pose and finger targets, converted into joint targets by damped least-squares inverse kinematics.Source: A. Simulation Settings · pp. 4–5
- PartManip contains 363/63 training/validation doors and 200/40 drawers, while ArticuRiddle has 100 derived objects with altered handles or modified joint axes that preserve appearance.The two ArticuRiddle edit types form visually similar pairs with different kinematics.Source: A. Simulation Settings · p. 5, Figure 4 · p. 5
- Success requires opening a door beyond 30° or a drawer beyond 0.2 m; articulation estimation is assessed with joint-type accuracy and ASRx for x = 10 or 20.Revolute ASRx also requires pivot error below x cm, while prismatic predictions are evaluated by the axis-angle criterion.Source: A. Simulation Settings · p. 5
- The comparison includes explicit estimate-then-plan systems and learned manipulation policies, with H-SAUR selecting interactions by predicted displacement and Wang et al. replanning from observed part motion.PartManip learns from point clouds using an expert policy distilled through DAgger.Source: A. Simulation Settings · p. 5
- 61.7% success on ArticuRiddle exceeds PartManip’s 32.4%, and the method outperforms AKM and H-SAUR in joint-type accuracy and ASR10/20 across all splits.The method also performs best throughout the reported comparisons, including training and validation doors and drawers.Source: A. Simulation Settings · p. 5
- On training doors, success rises from 37.4% with Wang et al. to 68.2%, and on training drawers from 44.2% to 89.0%.Against PartManip, the method's gains are 23.7/10.7 points on door train/validation and 21.9/21.1 points on drawer train/validation.Source: A. Simulation Settings · p. 5
B. Articulation Representation
The paper compares posterior-derived articulation flow with observed displacement, a frozen prior, and a learned latent representation, finding that explicit flow generalizes better to ambiguous objects.
B. Articulation Representation
- The representation comparison holds policy architecture, reward, and training budget constant across alternatives.This controls the comparison while testing how belief information is exposed to the point-cloud policy.Source: B. Articulation Representation · p. 5
- Observed displacement is unavailable before motion and unreliable for small single-step displacements, while the frozen NAP variant keeps prior hypotheses and weights fixed throughout an episode.Observed flow contains only the latest action and can be dominated by tracking noise.Source: B. Articulation Representation · p. 5
- The latent alternative maps privileged articulation state to a 16-dimensional code and predicts it from a 10-step history, but generalizes poorly because it must learn how each contact point relates to the axis.The policy receives the privileged code during training and the predicted code at evaluation.Source: B. Articulation Representation · pp. 5–6
- Posterior updating improves door train, door validation, and ArticuRiddle success by 6.1, 9.9, and 13.7 points over NAP without posterior updates.The latent variant reaches 36.2% on ArticuRiddle, where the flow's explicit contact-point/axis relation generalizes across geometry.Source: Table II · p. 6, B. Articulation Representation · p. 6
C. Ablation Studies
The ablations compare task displacement, KL divergence, and entropy decrease as interaction rewards; entropy decrease performs best across the reported manipulation-success columns. Its signed return also aligns better with final axis accuracy than KL divergence, which can reward belief changes that do not provide articulation information.
C. Ablation Studies
- The ablation compares joint displacement, KL divergence, and entropy decrease as interaction rewards, and also evaluates training solely through distillation without reinforcement learning.Information rewards are standardized; their weighting is swept over λi ∈ {0.05, 0.1, 0.2}, with the best setting reported for each.Source: C. Ablation Studies · p. 6
- Entropy decrease performs best in every manipulation-success column, while the criteria are accumulated over episodes rather than used only to select the next action.The belief entropy also reflects joint state, the observation process, and the particle filter.Source: C. Ablation Studies · p. 6
- Figure 5 ranks episodes by accumulated reward and compares their final axis-error fractions; it also plots median opening against the success threshold for each reward-trained policy.The caption reports bootstrapped 95% confidence intervals for episodes in the top plot and assets in the bottom plot.Source: Figure 5 · p. 6
- For entropy-return episodes, the fraction of doors ending above 10° axis error falls from 0.65 in the lowest-scoring quarter to 0.09 in the highest.By contrast, for KL-return episodes, that fraction rises from 0.27 to 0.94; entropy decrease can offset temporary belief-entropy rises, while KL divergence rewards large belief changes even when uninformative.Source: C. Ablation Studies · p. 6
D. Real-World Evaluation
The real-world evaluation transfers a simulation-trained system to unseen articulated objects without real-world fine-tuning. Interaction reduces axis errors and achieves manipulation success in both revolute and prismatic trials, while failures include grasping and visual-tracking problems.
D. Real-World Evaluation
- The system is retrained in simulation with a 6-DoF YAM arm and deployed without real-world fine-tuning; belief updates take 12 ms per step.The hardware setup uses a parallel-jaw gripper and RealSense D455 camera, with simulated camera pose, calibration bias, and depth noise randomized.Source: D. Real-World Evaluation · pp. 6–7
- Interaction reduces initial axis errors to 1.7–3.6°, with 65% success on revolute trials and 85% on prismatic trials.Table IV defines success as opening a door by more than 30° or a drawer by more than 20% of its opening range.Source: D. Real-World Evaluation · p. 7, Table IV · p. 7
- The prior is obtained by running NAP on a mesh extruded from one closed-state RGB-D frame, without annotation or interaction.The evaluation tests unseen revolute and prismatic objects varying in size, appearance, handle configuration, and hinge side, with 10 trials per object.Source: D. Real-World Evaluation · p. 7
- Six failures arise from the gripper not closing firmly on the handle, and four from mask leakage or late-episode self-occlusion during opening.The point-cloud visualizations show the articulation belief progressively converging toward the true joint as observed part motion drives the flow field.Source: D. Real-World Evaluation · p. 7, Figure 6 · p. 7
V. CONCLUSION
The conclusion presents a closed-loop framework for reducing uncertainty about articulated-object kinematics through interaction. It reports improved manipulation and articulation estimates, zero-shot transfer to four real-world objects, and boundaries tied to the supported joint set, grasping, and observable motion.
V. CONCLUSION
- The framework combines a particle-based articulation belief initialized from a generative prior, Bayesian updates from observed part motion, a flow-field policy representation, and an information reward with task reward.The policy is intended to gather useful information while manipulating.Source: §V · p. 7
- Experiments show improved manipulation over previous work, with interaction increasing joint-type accuracy and reducing axis and pivot errors.The policy also transfers zero-shot to four real-world objects while refining its articulation estimate and flow.Source: §V · p. 7
- The current method considers single-DoF joints covered by its initial hypothesis set and relies on a handle to grasp the part.The authors identify more complex, multi-part mechanisms as a direction for future work.Source: §V · p. 7
- Performance is limited when the robot fails to induce observable motion or when tracking is unreliable.Source: §V · p. 7