Robot learning papers put deployment details under test

DexSim2Real targets contact-rich hand manipulation without real demonstrations for the sim-to-real methods. It uses GPT-4V as a visual realism critic, then optimizes simulation randomization over lighting, texture, friction, mass, and camera noise. The policy also combines RGB, tactile readings, and proprioception through cross-attention during contact.

The reported hardware result is strong for this corpus: 78.2% average success over six real-world tasks on a Franka Panda with an Allegro Hand. The paper also reports an 8.3% average sim-to-real gap, compared with 28.5% for vanilla domain randomization and 19.2% for active domain randomization. The largest value is the mechanism, since the visual critic gives randomization a measurable target instead of relying only on hand-set ranges.

Anticipation-VLA addresses long tasks by keeping a stack of active goals and reachable subgoals. A high-level model proposes text and image subgoals. A value model then checks whether the task is achieved, improving, or stalled, and that status controls whether the system continues, refines the subgoal, or pops a completed goal.

The paper reports 63.2 success on Libero-Long, ahead of 54.6 for the underlying pi_0.5-style policy and 53.2 for a VLM-assisted version. In real-world Arx-X5 tests, the excerpt gives relative gains rather than exact success rates: +60% in seen configurations and +107% in unseen configurations. The useful lesson is operational. Long-horizon VLA execution needs state-dependent subgoals and progress checks, not a fixed recipe set before the rollout starts.

Two papers treat data collection and robot cost as core research constraints. Phone2Act turns an Android phone into a 6-DoF teleoperator, sends pose events through ROS 2, and records synchronized robot data directly in LeRobot format. Fine-tuning GR00T-N1.5-3B on 130 collected episodes produced 9 successes in 10 Dobot CR5 ball-to-basket trials, with 350–440 ms end-to-end latency.

VILAS builds a roughly $8,000 manipulation setup using a Fairino FR5 arm, dual RealSense cameras, teleoperation, and a kirigami soft gripper extension for fragile objects. On grape grasping, pi_0.5 reached 84% single-grasp success, while GR00T N1.6 reached 58% multi-grasp success and the lowest mean inference latency among the tested models. These results make cost, logging format, and gripper mechanics part of the VLA performance claim.

The world-model papers focus on what planning states preserve and how planners fail under triggers. Latent State Design proposes judging a world model by the function its latent state must support: prediction, control, planning, memory, grounding, or counterfactual reasoning. It offers a six-role taxonomy and a seven-axis evaluation matrix, but reports no new benchmark score.

TRAP gives the security counterpart. It attacks world-model planners by changing the ranking of imagined trajectories with a deployment-time visual patch. On DreamerV3 Crafter, it reports 98.1% attack success rate and a 63.2% mean return drop. On several DreamerV3 Atari and TD-MPC2 DMControl tasks, it reports 100% attack success. Together, the papers point to a concrete requirement for planned-control agents: evaluate the latent state for the task, then test whether the planner’s trajectory ranking can be manipulated.