Close Menu
  • AI
  • Content Creation
  • Tech
  • Robotics
AI-trends.todayAI-trends.today
  • AI
  • Content Creation
  • Tech
  • Robotics
Trending
  • OpenRouter Cuts Three Suppliers After AI Immediate Publicity Findings – Unite.AI
  • A Group of Researchers from Columbia and CertiK Introduce TensorLean for Formal Verification of GPU Kernels
  • Collectively AI Discloses Immediate Leak After Coding Agent Uncovered Credential – Unite.AI
  • Dragos Finds Immediate Leaks and Stolen Keys in AI Provide Chains
  • Lean and Mathlib Communities Certify Claude-Assisted Riemann Proof – Unite.AI
  • AWS Routes HyperPod Inference Throughout Areas From One Endpoint – Unite.AI
  • DailyAI Briefing: A Lean-checked Riemann proof, plus 4 extra
  • Anthropic Shares Drop 8.5% in Second Session Throughout AI Selloff – Unite.AI
AI-trends.todayAI-trends.today
Home»Tech»The Adaptive Diffusion Policy in Real-World Robotics: Latent-Space-Reinforcement-Learning (DSRL) Approach

The Adaptive Diffusion Policy in Real-World Robotics: Latent-Space-Reinforcement-Learning (DSRL) Approach

Tech By Gavin Wallace30/06/20254 Mins Read
Facebook Twitter LinkedIn Email
This AI Paper Introduces Differentiable MCMC Layers: A New AI
This AI Paper Introduces Differentiable MCMC Layers: A New AI
Share
Facebook Twitter LinkedIn Email

Learning-Based Robotics

The use of data-driven algorithms to replace manual coding has made huge progress in the development of robotic control systems. Modern robots do not rely on programming but instead learn from observing and imitating actions. This type of learning is often based on behavioral cloning and allows robots to operate effectively in structured environments. Transferring these behaviors to dynamic real-life scenarios is a difficult task. Robots are not just required to perform the same actions, but must also be able to adjust and improve their response when faced with unfamiliar environments or tasks. This is crucial for achieving generalized behavior.

The Challenges of Traditional Behavioral Cloning

The dependence of robot policy learning on human demonstrations is one of its main limitations. They are then used for supervised-learning to generate initial policies. Additional demonstrations may be required to train them if these policies are unable to generalize in a new setting or do so accurately. This inefficiency is due to the inability of robots to adapt policies to their own experience. The use of reinforcement learning to improve autonomously is possible, but its inefficiency with samples and dependence on accessing complex policy models makes it inappropriate for most real-world applications.

There are limitations to the current diffusion-RL integration

To refine robot behavior, different methods have been tried. Some of these efforts focused on the modification of early stages in the diffusion process, or the application additive adjustments to the policy outputs. Some have attempted to maximize actions through evaluating the expected rewards in denoising. Although these approaches improved the results of simulations, they required extensive computations and direct access into policy parameters. They are not practical for models that use black boxes or proprietary algorithms. On top of that, these models struggle with instability caused by backpropagation through diffusion chains.

The DSRL Framework for Latent Noise Policy Optimization

Researchers at UC Berkeley (UCB), the University of Washington and Amazon developed a method called Diffusion Steering Via Reinforcement-Learning (DSRL). The adaptation is shifted from changing the weighting of the policies to optimizing latent noise in the diffusion models. DSRL does not generate actions using a Gaussian distribution but instead trains a second policy which selects input noise to steer the resulting action towards desired outcomes. It is possible to use reinforcement learning for fine-tuning behaviors, without having to alter the base model.

The Latent Noise Space and Decoupling of Policy

Researchers reconfigured the learning environment through mapping of the initial action space into a space with latent-noise. This transformed set-up allows actions to be selected by selecting the latent sound that produces them, through diffusion policies. DSRL uses only the forward-outputs of the policy to create a reinforcement learning system that is completely independent from the base policy. It is adaptable for real robotic systems that only have black-box accessibility. This policy can then be trained by using actor-critic techniques, saving the cost associated with backpropagation. It allows online learning via real-time interactivity and offline learning using pre-collected datasets.

The Empirical Benefits and the Practical Results

The method proposed showed significant improvements in data and performance efficiency. DSRL, for example, improved the task completion rate from 20% to 90% in a real-world robot task within less than 50 online interactions. The performance of the robot has increased by more than 4 times with only minimal information. The method was also tested on a generalist robotic policy named π₀, and DSRL was able to effectively enhance its deployment behavior. All of these results were achieved with no modification to the diffusion policy, or any access to its parameters. They demonstrate the method’s usefulness in limited environments such as API only deployments.

The conclusion of the article is:

Summary: The research has addressed the key issue of policy adaptation in robotics without extensive retraining, or accessing models directly. Incorporating a steering mechanism based on latent-noise, the team has developed an effective tool for robot learning in real world situations. The strength of the method lies in its stability, efficiency and compatibility, which makes it an important step forward for adaptable robot systems.


Take a look at the Paper You can also find out more about the following: Project Page. The researchers are the sole owners of all credit. Also, feel free to follow us on Twitter Don’t forget about our 100k+ ML SubReddit Subscribe Now our Newsletter.


Nikhil has been an intern consultant with Marktechpost. The Indian Institute of Technology in Kharagpur offers him a dual degree integrated with Materials. Nikhil, an AI/ML fanatic, is constantly researching AI/ML applications for biomaterials and other biomedical fields. Material Science is his background. His passion for exploring and contributing to new advances stems from this.

ada Policy Robotics space
Share. Facebook Twitter LinkedIn Email
Avatar
Gavin Wallace

Related Posts

A Group of Researchers from Columbia and CertiK Introduce TensorLean for Formal Verification of GPU Kernels

22/11/2026

OpenAI Ships GPT-Stay-1.1 for Voice Brokers, Cuts Idle Value to $0.02/Min

21/11/2026

Meet Pondering Machines Inkling 2 Mini: 26B Looped MoE Runs on a 32GB Mac

21/11/2026

Baseten Releases Bison-1 and Groundhog-1, Topping Choice Mannequin Index

21/11/2026
Top News

OpenAI Updates GPT-5 Following Users’ Revolt

AI Models for War – What they Look Like

Gemini 3 pro: I’m the Next leap in Intelligence

Tips on how to Declare Your Reduce of Apple’s $250 Million Siri Settlement

The Pentagon has been unable to blacklist anthropomorphic images after a judge blocked their attempt

Load More
AI-Trends.Today

Your daily source of AI news and trends. Stay up to date with everything AI and automation!

X (Twitter) Instagram
Top Insights

Google AI Launches Multi Agent System Search: MASS is a New AI Agent Optimizing Framework to Improve Prompts, Topologies, and Topologies.

08/06/2025

Biomni R0: Agentic LLMs are trained end-to-end with multi-turn reinforcement learning for expert-level intelligence in biomedical research

06/09/2025
Latest News

OpenRouter Cuts Three Suppliers After AI Immediate Publicity Findings – Unite.AI

22/11/2026

A Group of Researchers from Columbia and CertiK Introduce TensorLean for Formal Verification of GPU Kernels

22/11/2026
X (Twitter) Instagram
  • Privacy Policy
  • Contact Us
  • Terms and Conditions
© 2026 AI-Trends.Today

Type above and press Enter to search. Press Esc to cancel.