New robotics and physical-AI papers, with what each one means for the industry.
Reinforcement Learning (RL) substantially improves planning performance in autonomous-driving Vision-Language-Action (VLA) models. However, existing RL methods rely on scalar trajectory rewards that measure overall quality but fail to identify specific planning errors or provide actionable correction signals.
Existing real-world reinforcement learning methods improve robot policies by combining human corrections with autonomous rollouts. However, accumulated experience varies in its value for policy improvement, and even successful trajectories can contain inefficient or redundant behavior, motivating more selective experience reuse.
Generalist robot policies are trained with imitation learning and are bounded by the quality and coverage of their demonstrations. Reinforcement learning can improve and extend them, either by fine-tuning the policy weights or by training a small module to steer the input noise of a frozen generative policy.
Multi-agent reinforcement learning (MARL) has become a prominent research direction, particularly within collective robotics, where the ARGoS simulator is widely used to study emergent swarm behavior. Yet despite this growth, systematic support for reuse and reproducibility in MARL-based swarm robotics remains limited and insufficiently standardized: a single experiment is assembled from manually aligned simulator files, training scripts, configuration values, paths, seeds, and environment details, so reuse and reproduction stay fragile even within one research group.
UniIntervene++ is a preprint describing an online learned intervention agent that selects among the robot's own RL policy, trajectory correction, and a task-structured CodePolicy during real-robot reinforcement learning. Across five manipulation tasks on a UR7e arm it reaches 89.67% average success, 6 points above the best baseline, while cutting human-controlled steps to 0.77%. This could reduce the skilled-operator burden needed to make online robot learning practical.
This preprint introduces BVER, a bidirectional curriculum that grows start states from the goal and goals from the initial state distribution, biased toward unexplored task space and steered together RRT-style. On quadruped box climbing it reaches 95% success on a 0.4 m box in about 65% fewer iterations than the best reference-free baseline and is the only reference-free method tested to climb a 0.7 m box. It matters because it removes the need for demonstrations or hand-designed rewards in sparse-reward robot learning.
This preprint introduces VIGOR, a model-based reinforcement learning framework that adds latent-space consistency to TD-MPC2 for zero-shot generalization to unseen visual distractions. On six DeepMind Control Suite tasks and four Robosuite tasks, it improves mean zero-shot returns by 3.4% over the second-best baseline on DMC and by 43.6% on Robosuite while matching TD-MPC2's sample efficiency. The work matters because it addresses a two-level compounding failure unique to recursive latent dynamics, which is critical for deploying sample-efficient MBRL in visually varying environments.
In this preprint, researchers introduce Adaptive Mask Selection and Composition (AMSC), a lifelong reinforcement learning method that uses online task-similarity embeddings to choose and weight previously learned policy masks when adapting to new tasks. On CT28 and MiniGrid benchmarks, it reaches AUC/forward transfer of 0.94/0.94 and 0.73/0.54 with zero measured forgetting, exceeding modular baselines. The approach matters because it enables selective policy reuse instead of reusing an entire history, reducing interference and helping agents accumulate skills without destructive forgetting.
This preprint introduces SCOUT, an offline multi-agent reinforcement learning framework that pairs a per-agent flow-matching behavior prior with a decomposed value function and applies test-time Stein variational refinement to action candidates. Across SMACv1 and MA-MuJoCo, it reports the best average scores among all listed baselines (SMACv1 average reward 16.1 vs 14.9 for MAC-Flow; MA-MuJoCo normalized score 82.12 vs 76.37 for MAC-Flow) and improves on all reported offline-to-online transitions. The result matters because it shifts the coordination/value trade-off from a fixed training-time regularization to an adjustable inference-time compute budget.
In this preprint, PROMO conditions a single quadruped locomotion policy on a three-way semantic preference over tracking, stability, and energy efficiency. On a Unitree Go2, changing only the preference reduces specific energy by up to 30.4%, position error by 38.7%, and peak body-attitude deviation by 59.0% relative to a balanced setting. This gives legged robots an operator-facing trade-off dial without retraining or switching controllers.
In this preprint, the authors present ReCo, a framework that trains a quadruped locomotion policy to have a predictable command response and then couples it with policy-aware MPC to jointly plan base motion and a six-DoF arm for continuous manipulation. In simulation, ReCo cuts end-effector position RMSE by 28.7% and orientation RMSE by 27.4% versus the best baselines and succeeds in 75/100 runs versus 42/100 for the strongest baseline; the full stack runs onboard a Unitree Go2 with an ARX X5 arm. This matters because it makes continuous world-frame manipulation while walking practical without a full accurate contact model.
This is a preprint describing ColoACT, an imitation-learning navigation system for a compact self-propelled endoscopic robot. The policy fuses RGB, relative depth, and a gradient-based pseudo-elevation map with an Action Chunking Transformer and temporal ensembling. In ex-vivo porcine colons it reached 85.4% success on straight and 72.5% on curved segments, with smooth steering and success on 90-degree turns (70%) and double bends (60%), suggesting a path toward reducing physician burden in colonoscopy.
This arXiv preprint introduces FERPO, an on-policy maximum-entropy reinforcement-learning algorithm that updates a policy from critic values without taking critic action derivatives. It roughly matches REPPO on DMC and ManiSkill while reaching high scores with fewer environment interactions, and at K=16 its uncached actor-critic updates are 6.9% faster than REPPO, or 22.2% faster with cached actor updates. The method matters for massively parallel robot-policy training because it avoids unreliable learned action gradients and improves wall-clock and sample efficiency.
This preprint introduces MFPG-PPO, an algorithm that uses abundant low-fidelity simulator data as a control variate to reduce variance in scarce high-fidelity policy-gradient updates. On the hardest simulated locomotion task it matches PPO trained with 16 times more high-fidelity data, and on a physical Franka arm it reaches 91.7% success with only 4 real episodes per update, versus 55.0% for PPO on high-fidelity data alone. If validated, the approach could make on-policy robot learning much more sample-efficient in real-world deployments.
This preprint presents TacEx, a framework that directs reinforcement-learning exploration by weighting tactile predictive uncertainty instead of total state uncertainty. In simulation, tactile-driven exploration produces contact-rich examples, emergent grasping without task rewards, and improved offline pick-and-place and frozen VLA post-training. It matters because it makes manipulation data collection and VLA adaptation far more sample-efficient by focusing learning where contact occurs.
This preprint introduces GroundingPI, a 4B grounding foundation model that predicts object points and boxes as quantized token coordinates rather than using a general-purpose vision-language backbone. Across 34 grounding benchmarks it averages 73.68%, ahead of a larger GPT-6 Astra baseline at 71.54%, and as a visual backbone it improves downstream manipulation (RoboTwin 2.0, RoboCasa-GR1) and autonomous driving (nuScenes L2 0.296 m). The result matters because it argues grounding is a distinct perceptual layer that can make embodied foundation models more precise.
This preprint introduces AC-PAC-NMPC, a hybrid controller that places RL-trained actor, critic, and sensor-prediction models inside a sampling-based stochastic NMPC framework to provide finite-time probabilistic safety bounds and long-horizon vision-based navigation. On a fixed-wing UAV with an onboard depth camera, it raised real-robot success from 40% (both RL actor and best PAC-NMPC baseline) to 80% with the lowest cost; in simulation it reached 90% success versus 85% for the strongest baseline. This matters for agile robots that need learned long-range behavior without sacrificing formal collision-probability guarantees.
This preprint introduces a battery-aware reinforcement-learning controller for aggressive quadrotor flight. On a 38 g Crazyflie Brushless, a policy with increased thrust-command authority plus a filtered voltage input cuts 3.84 m/s circle tracking error by 49.3% versus stock-authority RL and reduces 20-lap race time from 106.22 s to 95.24 s. The work shows that explicit voltage feedback can complement existing flight-controller compensation, offering a practical route to better agility on battery-powered drones.
Constrained Reinforcement Learning has recently gained increasing attention in the field of Safe Autonomous Driving, where the general mechanism is to maximize the expected reward while keeping the overall action risk bounded. In this way, the safety issues arising in AD can be mitigated through constrained actions.
RL policies trained in GPU simulation transfer zero-shot to a commodity humanoid on rough terrain.
Papers come from arXiv robotics feeds. Where our summary is not written yet, you see the opening lines of the abstract.