Research and policy news on artificial intelligence and robotics, updated three times a day from the public feeds of research laboratories, universities and EU institutions, with a daily selection of new research papers. Each item has a short summary and a link to the source.
MIT and New York University researchers introduce FLEX, a feature-editing pipeline that lets people steer robot motion planning through conversational text. Using a large language model, the robot is meant to infer the broader intent behind an instruction, such as avoiding a laptop while carrying coffee. The team reports tests in simulated scenarios and will present the study at the Conference on Robot Learning.
The authors examine what happens when sampling-based planners such as MPPI search hard against reward models learned from human preferences. They report that stronger search can cut true return by up to 76% while held-out label accuracy stays high, and they derive a regularized temperature rule that they find safest among tested fixes.
The authors test whether OpenVLA obeys negated instructions on LIBERO-Goal and probe where it encodes its executed goal using patching, probes and steering. They report that negation has no detectable effect, with the forbidden task executed in 79% of rollouts, and that goal edits only work early in the network or across every action token.
The authors build ABHBench, a multimodal benchmark drawn from the game Detroit: Become Human, to assess how foundation models choose when acting as android agents in moral dilemmas, scoring harmlessness, obedience and self-preservation. Across 12 models, they report that 75% shift significantly between text-only and situated multimodal versions.
The authors propose a counterfactual audit for grasp-feasibility predictors that places an obstacle in the approach corridor and has a motion planner relabel the grasp as infeasible. Across eight predictors, they report pair AUROC falling to chance, and note that a clearance-aware corridor reading recovers performance only for trained obstacle placements.
The authors frame the decision to verify or steer an agent mid-trajectory as a causal inference problem and build a framework for estimating dynamic effects of randomized interventions from a modest set of trajectories. They describe using the estimated parameters to build safety leaderboards for audit prioritization, applying the method to tau-bench.
The authors run a controlled simulation study of coding agents that write and refine robot policies, varying model capability, skill priors, feedback and curriculum across nine manipulation tasks and 28K rollouts. They report that the main bottleneck tends to be turning available information into correct policy updates rather than accessing that information.
The authors survey 289 U.S. adults on 45 scenarios across nine life domains to learn which tasks people would hand to AI agents. They report that trust, desired AI leadership and expected savings predict greater automation preference, while concern about lost personal fulfillment predicts less, with respondents favoring retained human oversight.
The authors introduce STRIDE, which trains humanoid walking with a reward based on following commanded accelerations measured by the onboard IMU, combined with off-policy learning, high joint gains and symmetry augmentation. In simulation, they report Unitree G1 and Booster T1 walking within 14 minutes of samples and 81% fewer training falls.
The authors report that AMP1, a harness letting a generalist AI model directly command single-arm, bimanual, and mobile manipulators, scores 85.6% zero-shot on RoboLab-120 versus 49.2% and 42.9% for top VLA and WAM policies, without demonstrations or task training. They also describe 104 successes in 108 real trials and say adding cross-episode memory cut RoboDojo bimanual failures 4.6x.
AI-generated summaries. Items, titles and links come from the publishers' public news feeds and from arXiv, and each title links to the original. Summaries are written and checked automatically from the source text and are not individually reviewed by a person; the original is authoritative. Research papers are preprints, not peer reviewed. Llull Lab is independent of the publishers listed. How it is made.