| phenomenon | visuomotor policy learning— “In this section, we define the visuomotor policy learning problem and present an overview of our approach.” |
| claim-type | method— “To this end, we develop a method that can be used to learn policies that map raw image observations directly to torques at the robot’s motors.” |
| method — evidence | — “We address these challenges by developing a guided policy search algorithm for sensorimotor deep learning, as well as a novel CNN architecture designed for robotic control.” |
| transfer | from artificial intelligence (deep learning / computer vision): deep convolutional networks trained end-to-end from perception to control for robotic manipulation— “as well as a novel CNN architecture designed for robotic control” | transfers[0] |
| mechanism | guided policy search: convert reinforcement-learning policy search into supervised learning of a deep network, with a trajectory-optimizer acting as the teacher— “guided policy search method, which transforms policy search into supervised learning” | mechanisms[0] |
| builds-on | convolutional neural networks (deep learning)— “We represent our policies with convolutional neural networks” | claim_relations[0] |
| builds-on | reinforcement learning— “supervision provided by a simple trajectory-centric reinforcement learning method” | claim_relations[1] |
| builds-on | Bregman ADMM (BADMM) (Wang and Banerjee, 2014)— “We show that this can be formalized as an instance of Bregman ADMM (BADMM) (Wang and Banerjee, 2014), which can be used to show that the algorithm converges to a locally optimal solution.” | claim_relations[2] |
| extends | Levine and Abbeel, 2014; Levine et al., 2015 (the authors' own conference papers)— “Some of the material in this article has previously appeared in two conference papers (Levine and Abbeel, 2014; Levine et al., 2015), which we extend to introduce visual input into the policy.” | claim_relations[3] |
| extends | Mordatch and Todorov (2014)— “The use of ADMM for guided policy search was also proposed by Mordatch and Todorov (2014) for deterministic policies under known dynamics.” | claim_relations[4] |
| sense | architecture quality - MEASURED ON A PROXY: accuracy on the pose-estimation pretraining task, chosen to isolate the architectures from other confounding factors— “To isolate the architectures from other confounding factors, we measure their accuracy on the pose estimation pretraining task described in Section 5.2. This is a reasonable proxy for evaluating how well the network can overcome two major challenges in visuomotor learning” | senses[0] |
| sense | failure at insertion - defined by a threshold taken from the object itself: since the peg is 0.5 units long, a larger distance means the controller cannot insert— “Since the peg is 0.5 units long, distances above this amount correspond to controllers that cannot perform an insertion.” | senses[1] |
| sense | observable - split by phase, and the split is the method's central assumption: the full system state is observable AT TRAINING TIME BUT NOT AT TEST TIME, which is what lets a trajectory-centric RL algorithm supply supervision the deployed policy never gets— “In our method, the full state of the system is observable at training time, but not at test time. For most tasks, providing the full state simply requires positioning objects in one of several known positions for each trial during training” | senses[2] |
| control | two baselines that pre-train the vision layers for pose prediction instead of training end-to-end, one with the SAME architecture as the full policy — held— “Our results demonstrate improvements in consistency and generalization from training visuomotor policies end-to-end, when compared to training the vision and control components separately.” — rules out that the improvement comes from the network architecture rather than from training the vision and control layers jointly end-to-end— “We compared to two baselines, both of which train the vision layers in advance for pose prediction, instead of training the entire policy end-to-end.” | controls[0] |
| audit | guided policy search —operationalised-as→ supervised learning — Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but the span says guided policy search TRANSFORMS policy search into a supervised learning problem; operationalised-as is loose.— “Guided policy search transforms policy search into a supervised learning problem, where the training set is generated by a simple trajectory-centric RL algorithm” | concept_edges[0] |
| note | guided policy search —operationalised-as→ supervised learning — THE METHOD'S WHOLE MOVE IS A CONVERSION BETWEEN PROBLEM CLASSES, stated as such: '…search TRANSFORMS policy search INTO A SUPERVISED LEARNING…' The hard problem is not solved; it is turned into an easier one whose labels another algorithm supplies.— “Guided policy search transforms policy search into a supervised learning problem, where the training set is generated by a simple trajectory-centric RL algorithm” | concept_edges[0] |
| weight | guided policy search —operationalised-as→ supervised learning — 0.9— “Guided policy search transforms policy search into a supervised learning problem, where the training set is generated by a simple trajectory-centric RL algorithm” | concept_edges[0] |
| outcome | control — held— “We compared to two baselines, both of which train the vision layers in advance for pose prediction, instead of training the entire policy end-to-end.”— “Our results demonstrate improvements in consistency and generalization from training visuomotor policies end-to-end, when compared to training the vision and control components separately.” | controls[0] |
| for which use | limitation — the training procedure as described: it requires TO POSITION THE OBJECTS IN THE SCENE AT CONSISTENT POSITIONS— “to position the objects in the scene at consistent positions” | limitations[0] |
| for which use | limitation — the generalisation claim: OUR CURRENT METHOD DOES NOT GENERALIZE TO DRAMATICALLY DIFFERENT SETTINGS, ESPECIALLY WHEN VISUAL DISTRACTO[rs occlude]— “Although we demonstrate moderate generalization over variations in the scene, our current method does not generalize to dramatically different settings, especially when visual distractors occlude the manipulated object or break up its silhouette in ways that differ from the training.” | limitations[1] |
| for which use | limitation — the range of tasks the method can address: THE REQUIREMENT TO OBSERVE THE FULL STATE DURING TRAINING LIMITS THE TASKS TO WHICH THE METHOD CAN BE APPLIED— “However, the requirement to observe the full state during training limits the tasks to which the method can be applied.” | limitations[2] |
| for which use | limitation — any extension to freely moving objects: those TASKS ... REQUIRE MORE EXTENSIVE INSTRUMENTATION, SUCH AS MOTION CAPTURE— “However, tasks that require, for example, manipulating freely moving objects require more extensive instrumentation, such as motion capture.” | limitations[3] |
| for which use | limitation — every result produced through the visual encoder: THE FEATURE POINT REPRESENTATION ... ASSUMES THAT THE LEARNED FEATURES ARE PRESENT AT ALL TIMES, AND ONLY ONE INSTANCE OF EACH FEATURE IS EVER PRESENT IN THE IMAGE— “The feature point representation is very simple, since it assumes that the learned features are present at all times, and only one instance of each feature is ever present in the image.” | limitations[4] |
| for which use | limitation — applying this method to a new task: the full state must be observable during training, which the authors call minor when objects can be positioned consistently and not minor for tasks manipulating freely moving objects— “However, the requirement to observe the full state during training limits the tasks to which the method can be applied. In many cases, this limitation is minor, and the only “instrumentation” required at training is to position the objects in the scene at consistent positions. However, tasks that require, for example, manipulating freely… | limitations[5] |
| for which use | limitation — the architecture comparison in this paper: it is measured on the pose-estimation pretraining task, described by the authors as a reasonable proxy for the visuomotor challenges— “To isolate the architectures from other confounding factors, we measure their accuracy on the pose estimation pretraining task described in Section 5.2. This is a reasonable proxy for evaluating how well the network can overcome two major challenges in visuomotor learning” | limitations[6] |
| replication | untested |
| limitation | training requires positioning objects at consistent positions— “to position the objects in the scene at consistent positions” | limitations[0] |
| limitation | The method does not generalise to dramatically different visual settings, especially under occluding distractors.— “Although we demonstrate moderate generalization over variations in the scene, our current method does not generalize to dramatically different settings, especially when visual distractors occlude the manipulated object or break up its silhouette in ways that differ from the training.” | limitations[1] |
| limitation | The approach requires the full system state to be observable during training, which restricts the range of applicable tasks.— “However, the requirement to observe the full state during training limits the tasks to which the method can be applied.” | limitations[2] |
| limitation | Tasks with freely moving objects would need heavier instrumentation such as motion capture.— “However, tasks that require, for example, manipulating freely moving objects require more extensive instrumentation, such as motion capture.” | limitations[3] |
| limitation | The spatial feature-point representation makes simplifying assumptions about feature presence and uniqueness.— “The feature point representation is very simple, since it assumes that the learned features are present at all times, and only one instance of each feature is ever present in the image.” | limitations[4] |
| limitation | The method needs the full state observable during training, and the authors say exactly when that is minor and when it is not.— “However, the requirement to observe the full state during training limits the tasks to which the method can be applied. In many cases, this limitation is minor, and the only “instrumentation” required at training is to position the objects in the scene at consistent positions. However, tasks that require, for example, manipulating freely… | limitations[5] |
| limitation | The architecture comparison is not run on the target task but on a pretraining task, which the authors describe as a reasonable proxy.— “To isolate the architectures from other confounding factors, we measure their accuracy on the pose estimation pretraining task described in Section 5.2. This is a reasonable proxy for evaluating how well the network can overcome two major challenges in visuomotor learning” | limitations[6] |
| limitation | The training-time observability the method depends on is bought with a physical constraint on the world, stated plainly: providing the full state 'simply requires positioning objects in one of several known positions for each trial during training'. The assumption is not about the algorithm but about how the laboratory is arranged.— “In our method, the full state of the system is observable at training time, but not at test time. For most tasks, providing the full state simply requires positioning objects in one of several known positions for each trial during training” | limitations[7] |