When designing reinforcement learning (RL) agents, a designer communicates the desired agent behavior through the definition of reward functions - numerical feedback given to the agent as reward or punishment for its actions. However, mapping desired behaviors to reward functions can be a difficult process, especially in complex environments such as autonomous racing. In this paper, we demonstrate how current foundation models can effectively search over a space of reward functions to produce desirable RL agents for the Gran Turismo 7 racing game, given only text-based instructions. Through a combination of LLM-based reward generation, VLM preference-based evaluation, and human feedback we demonstrate how our system can be used to produce racing agents competitive with GT Sophy, a champion-level RL racing agent, as well as generate novel behaviors, paving the way for practical automated reward design in real world applications.

Multi-Agent Reinforcement Learning (MARL) provides a powerful framework for learning coordination in multi-agent systems. However, applying MARL to robotics still remains challenging due to high-dimensional continuous joint action spaces, complex reward design, and non-stationary transitions inherent to decentralized settings. On the other hand, humans learn complex coordination through staged curricula, where long-horizon behaviors are progressively built upon simpler skills. Motivated by this, we propose CRAFT: Coaching Reinforcement learning Autonomously using Foundation models for multi-robot coordination Tasks, a framework that leverages the reasoning capabilities of foundation models to act as a "coach" for multi-robot coordination. CRAFT automatically decomposes long-horizon coordination tasks into sequences of subtasks using the planning capability of Large Language Models (LLMs). In what follows, CRAFT trains each subtask using reward functions generated by LLM, and refines them through a Vision Language Model (VLM)-guided reward-refinement loop. We evaluate CRAFT on multi-quadruped navigation and bimanual manipulation tasks, demonstrating its capability to learn complex coordination behaviors. In addition, we validate the multi-quadruped navigation policy in real hardware experiments.

Reinforcement Learning (RL) robot controllers usually aggregate many task objectives into one scalar reward. While large-scale proximal policy optimisation (PPO) has enabled impressive results such as robust robot locomotion in the real world, many tasks still require careful reward tuning and are brittle to local optima. Tuning cost and sub-optimality grow with the number of objectives, limiting scalability. Modelling reward vectors and their trade-offs can address these issues; however, multi-objective methods remain underused in RL for robotics because of computational cost and optimisation difficulty. In this work, we investigate the conflict between gradient contributions for each objective that emerge from scalarising the task objectives. In particular, we explicitly address the conflict between task-based rewards and terms that regularise the policy towards realistic behaviour. We propose GCR-PPO, a modification to actor-critic optimisation that decomposes the actor update into objective-wise gradients using a multi-headed critic and resolves conflicts based on the objective priority. Our methodology, GCR-PPO, is evaluated on the well-known IsaacLab manipulation and locomotion benchmarks and additional multi-objective modifications on two related tasks. We show superior scalability compared to parallel PPO (p = 0.04), without significant computational overhead. We also show higher performance with more conflicting tasks. GCR-PPO improves on large-scale PPO with an average improvement of 9.5%, with high-conflict tasks observing a greater improvement. The code is available at https://github.com/humphreymunn/GCR-PPO.

Figure 1: Text2Touch improves upon previous reward function design methods to increase the performance of robotic in-hand object rotation in rotation speed and grasp stability. We evaluate the performance of LLM-generated reward functions using only tactile and proprioceptive information in the real world. Designing reinforcement learning (RL) reward functions for dexterous in-hand manipulation remains a formidable challenge. Traditional approaches often rely on domain experts to painstakingly specify and tune reward terms [1], a process prone to suboptimal or unintended behaviours [2, 3]. Recent work has shown that large language models (LLMs) can generate policy or reward code for robotic tasks [4, 5, 6, 7, 8, 9], a notable step toward reducing manual engineering. However, these breakthroughs have primarily focused on conventional sensing modalities (vision, proprioception) for real-world validation [4, 10, 6, 7, 8]. To date, tactile sensing has not yet been integrated into automated reward generation via LLMs in either simulated or real-world settings. Vision based tactile sensing can provide detailed contact and force signals that visual sensing alone often fails to capture, especially under occlusions or subtle slip conditions [11].

Despite these successes, applying RL techniques to complex control problems remains a daunting undertaking, where initial attempts often result in underwhelming performance.

Reinforcement learning (RL) has achieved remarkable success in fields like robotics and autonomous driving, but adversarial attacks designed to mislead RL systems remain challenging. Existing approaches often rely on modifying the environment or policy, limiting their practicality. This paper proposes an adversarial attack method in which existing agents in the environment guide the target policy to output suboptimal actions without altering the environment. We propose a reward iteration optimization framework that leverages large language models (LLMs) to generate adversarial rewards explicitly tailored to the vulnerabilities of the target agent, thereby enhancing the effectiveness of inducing the target agent toward suboptimal decision-making. Additionally, a critical state identification algorithm is designed to pinpoint the target agent's most vulnerable states, where suboptimal behavior from the victim leads to significant degradation in overall performance. Experimental results in diverse environments demonstrate the superiority of our method over existing approaches.

Designing effective reward functions is a cornerstone of reinforcement learning (RL), yet it remains a challenging process due to the inefficiencies and inconsistencies inherent in conventional reward engineering methodologies. Recent advances have explored leveraging large language models (LLMs) to automate reward function design. However, their suboptimal performance in numerical optimization often yields unsatisfactory reward quality, while the evolutionary search paradigm demonstrates inefficient utilization of simulation resources, resulting in prohibitively lengthy design cycles with disproportionate computational overhead. To address these challenges, we propose the Uncertainty-aware Reward Design Process (URDP), a novel framework that integrates large language models to streamline reward function design and evaluation in RL environments. URDP quantifies candidate reward function uncertainty based on self-consistency analysis, enabling simulation-free identification of ineffective reward components while discovering novel reward components. Furthermore, we introduce uncertainty-aware Bayesian optimization (UABO), which incorporates uncertainty estimation to significantly enhance hyperparameter configuration efficiency. Finally, we construct a bi-level optimization architecture by decoupling the reward component optimization and the hyperparameter tuning. URDP orchestrates synergistic collaboration between the reward logic reasoning of the LLMs and the numerical optimization strengths of the Bayesian Optimization. We conduct a comprehensive evaluation of URDP across 35 diverse tasks spanning three benchmark environments. Our experimental results demonstrate that URDP not only generates higher-quality reward functions but also achieves significant improvements in the efficiency of automated reward design compared to existing approaches.

Vision-Language-Action (VLA) models have shown great potential in general robotic decision-making tasks via imitation learning. However, the variable quality of training data often constrains the performance of these models. On the other hand, offline Reinforcement Learning (RL) excels at learning robust policy models from mixed-quality data. In this paper, we introduce Reinforced robot GPT (ReinboT), a novel end-to-end VLA model that integrates the RL principle of maximizing cumulative reward. ReinboT achieves a deeper understanding of the data quality distribution by predicting dense returns that capture the nuances of manipulation tasks. The dense return prediction capability enables the robot to generate more robust decision-making actions, oriented towards maximizing future benefits. Extensive experiments show that ReinboT achieves state-of-the-art performance on the CALVIN mixed-quality dataset and exhibits superior few-shot learning and out-of-distribution generalization capabilities in real-world tasks.

Enabling a high-degree-of-freedom robot to learn specific skills is a challenging task due to the complexity of robotic dynamics. Reinforcement learning (RL) has emerged as a promising solution; however, addressing such problems requires the design of multiple reward functions to account for various constraints in robotic motion. Existing approaches typically sum all reward components indiscriminately to optimize the RL value function and policy. We argue that this uniform inclusion of all reward components in policy optimization is inefficient and limits the robot's learning performance. To address this, we propose an Automated Hybrid Reward Scheduling (AHRS) framework based on Large Language Models (LLMs). This paradigm dynamically adjusts the learning intensity of each reward component throughout the policy optimization process, enabling robots to acquire skills in a gradual and structured manner. Specifically, we design a multi-branch value network, where each branch corresponds to a distinct reward component. During policy optimization, each branch is assigned a weight that reflects its importance, and these weights are automatically computed based on rules designed by LLMs. The LLM generates a rule set in advance, derived from the task description, and during training, it selects a weight calculation rule from the library based on language prompts that evaluate the performance of each branch. Experimental results demonstrate that the AHRS method achieves an average 6.48% performance improvement across multiple high-degree-of-freedom robotic tasks.

Robotics Reinforcement Learning (RL) often relies on carefully engineered auxiliary rewards to supplement sparse primary learning objectives to compensate for the lack of large-scale, real-world, trial-and-error data. While these auxiliary rewards accelerate learning, they require significant engineering effort, may introduce human biases, and cannot adapt to the robot's evolving capabilities during training. In this paper, we introduce Reward Training Wheels (RTW), a teacher-student framework that automates auxiliary reward adaptation for robotics RL. To be specific, the RTW teacher dynamically adjusts auxiliary reward weights based on the student's evolving capabilities to determine which auxiliary reward aspects require more or less emphasis to improve the primary objective. We demonstrate RTW on two challenging robot tasks: navigation in highly constrained spaces and off-road vehicle mobility on vertically challenging terrain. In simulation, RTW outperforms expert-designed rewards by 2.35% in navigation success rate and improves off-road mobility performance by 122.62%, while achieving 35% and 3X faster training efficiency, respectively. Physical robot experiments further validate RTW's effectiveness, achieving a perfect success rate (5/5 trials vs. 2/5 for expert-designed rewards) and improving vehicle stability with up to 47.4% reduction in orientation angles.