Robot manipulation: learning, control and tactile feedback

Manipulation begins before contact with scene understanding and motion planning, then becomes a contact and control problem as the hand or gripper reaches the object. A complete system needs a defined observation, action, end effector, controller or policy, task objective, and evaluation protocol.

A model that succeeds at pick-and-place has not automatically solved insertion, deformable objects, tool use, or multi-finger dexterity. Task families create different sensing, control, and evaluation requirements.

PRISM reports 5,000+ trajectories across 25+ industrial tasks and multiple robot and teleoperation configurations, including insertion, packaging, installation, and conveyor sorting. Tactile observations cover only an unspecified subset, and the official dataset download remained pending when reviewed, so the paper is evidence of the announced collection design rather than an independently audited open benchmark.

A two-finger gripper can be reliable and easier to control for many industrial tasks. Multi-finger hands add contact options and human-tool compatibility but also increase sensing, calibration, action-space, control, and maintenance complexity. Whole-body manipulation adds balance, mobility, reach, and environmental contacts.

Compare systems by the task and embodiment they actually test, not by assuming that more degrees of freedom always produce better manipulation.

Modern manipulation research includes imitation learning, reinforcement learning, diffusion or flow-based policies, vision-language-action models, and predictive world or world-action models. These methods can share data and components while serving different roles.

Language can specify a task, vision can establish scene context, proprioception can expose robot configuration, and touch can ground the policy in physical contact. A world model may predict what follows an action, while a controller still needs to execute and correct the motion.

Tactile feedback is most defensible when the experiment isolates what changes after adding touch. A matched vision-only or no-touch baseline, synchronized inputs, latency reporting, disturbance tests, and closed-loop outcomes help show whether the contact pathway is doing useful work.

T-Rex investigates tactile-reactive VLA manipulation, while ReTouch investigates online-refined tactile prediction for contact-rich dexterity. Both are 2026 preprints with source-specific robots, datasets, baselines, and tasks. Their reported results should not be transferred to other systems without new evidence.

ADEPT provides a narrower matched example: one Flexiv-Sharpa square-and-round insertion condition reports 3/10 final success with vision only and 8/10 with visuo-tactile input, using ten physical trials per modality. That comparison is tied to one fixed-workbench hand, tactile representation, task, and training recipe; it is not a universal tactile-manipulation gain.

A credible comparison aligns the task, robot, end effector, observations, training data, action budget, controller rate, objects, environment, and evaluation procedure. It also states whether failures were retried, demonstrations were selected, or a person intervened.