A clearer view of contact-rich manipulation.

An independent, source-backed reference for robotic manipulation benchmarks and policies, spanning RGB, depth, tactile sensing and force feedback.

Two ways into the literature

Benchmarks describe what is tested: published tasks, evaluation domains, simulator versions, sensors, robot embodiments, demonstrations, metrics and protocols. Task figures show the suite or capability taxonomy whenever an appropriate author figure is available.

Policies describe how actions are produced: architecture, base model, deployment inputs, action outputs, model size, learning data, training resources and inference or control rates. Their figures show the method pipeline or architecture. Agent / RSI / ICL records also distinguish in-context examples, execution feedback, reusable code and memory, and parameter updates, with their reported adaptation and test protocols.

Supporting models provide tactile representations, estimated geometry or predicted futures. Their records preserve model inputs/outputs, pretraining and downstream scope without inventing action outputs or adding to the executable-policy count. Use the Record type filter to isolate models.

Surveys appear within relevant research categories, with their own review scope, taxonomy, literature snapshot and evaluation axes. They use conceptual figures and are counted separately from executable policies. Use the Record type filter to focus on surveys, models or executable policies.

A paper can contribute to both collections. It keeps one canonical research record, with collection-specific details and figures. Supporting datasets, tactile simulators and perception tests are labeled separately from closed-loop manipulation benchmarks.

Evidence before completeness claims

• Primary papers, author project pages, official repositories and model cards support the records. Secondary lists are discovery aids.
• Publication dates use the first public paper version where verified. An anonymous, undated project stays undated.
• Paper experiments and current software releases can use different simulators, datasets or checkpoints. Both scopes are retained.
• Undisclosed values stay Not reported. Unresolved values stay Not verified. Neither means zero, absent or closed source.
• Code, model weights and data have separate release and licensing fields. A project link does not establish an open release.

Physical observations and deployment inputs

Rendered ground-truth depth, reconstructed sensor depth, geometry predicted from RGB, tactile images, taxel force arrays and wrist force/torque are distinct observations. Privileged simulator state or tactile labels used during training are not automatically deployment inputs. Policy details describe these differences explicitly.

Model sizes identify the component when reported. A training GPU count, inference device, action throughput, environment step rate and closed-loop control frequency are different measurements. This atlas retains their source scope rather than turning them into a shared performance ranking.

Figures and attribution

Research figures come from author publications, projects or official repositories. Images are optimized locally and retain their source URL, author credit and original rights. Click a figure to enlarge it and follow its source. Missing figures are stated explicitly.

Keeping the atlas current

This is a reviewed snapshot dated 26 Oct 2026, not a live publication feed. The coverage page describes the search boundary and unresolved fields. The repository includes structured data, regeneration and validation scripts, and a review workflow for future additions.

Search accepts task names, simulators, model names, sensors and training details. Filters are collection-specific. Save papers in this browser, compare up to four records, export a filtered CSV, or share a URL that restores the current view.