How can robots acquire skills through interactions with the physical world? An interview with Jiaheng Hu

2026-02-12 · Source: ΑΙhub · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Advanced, short

Summary

Jiaheng Hu, Peter Stone, and Roberto Martín-Martín introduce SLAC (Simulation-Pretrained Latent Action Space for Whole-Body Real-World RL), a method addressing the challenges of real-world reinforcement learning (RL) for complex robotic systems like mobile manipulators. Traditional sim2real approaches are limited by the need for high-fidelity, task-specific simulations and difficulties in simulating contact-rich interactions or deformable objects. Real-world RL, while promising, faces issues of sample inefficiency and safety during exploration. SLAC tackles this by using a two-step process: first, it learns a safe and structured latent action space in a low-fidelity simulation via unsupervised RL. Second, this pretrained action space is used for real-world RL on downstream tasks, enabling sample-efficient and safe learning. The method achieved over 80% success rates on challenging visuo-motor, whole-body, and contact-rich tasks with a Tiago robot within an hour of real-world interaction, a significant improvement over previous methods.

Key takeaway

For AI Scientists developing real-world robotic control policies, SLAC offers a viable path to overcome the safety and sample-efficiency hurdles of direct real-world RL. You should consider integrating a low-fidelity simulation pretraining step to establish a safe and structured latent action space, enabling robust learning for complex, contact-rich tasks on high-degree-of-freedom robots like the Tiago. This approach can significantly reduce real-world interaction time and prevent robot damage.

Key insights

SLAC enables safe and sample-efficient real-world reinforcement learning for complex robots by pretraining a latent action space in low-fidelity simulation.

Principles

• Low-fidelity simulation can assist real-world RL.
• Pretraining a safe, structured action space mitigates real-world RL risks.

Method

SLAC learns a latent action space in simulation via unsupervised RL, designing objectives for safe and structured behaviors. This learned space then serves as the action space for real-world RL on downstream tasks.

In practice

• Apply SLAC for mobile manipulator skill acquisition.
• Use SLAC for tasks requiring whole-body motion and contact-rich interactions.

Topics

• Reinforcement Learning
• Robotics
• Sim2Real
• Latent Action Space
• Mobile Manipulators

Best for: AI Scientist, Research Scientist, AI Researcher, Robotics Engineer, Machine Learning Engineer

Related on AIssential

• True Positive Weekly #142 — Robotic foundation models leverage multimodal LLMs and specialized control modules for dexterous, generalizable robotic task execution.
• Pixels to Proofs: Probabilistically-Safe Latent World Model Control via Parallel Conformal Robust MPC — SLS^2 enables probabilistically-safe robot control from pixels by combining latent world models with conformal robust MPC.
• Video Friday: Digit Learns to Dance—Virtually Overnight — Advanced robotics are achieving new performance thresholds in control, learning, and human interaction.
• Learning of Robot Safety Policies via Adversarial Synthetic Scenarios — An adversarial gamification framework enables robust robot safety policy learning by iteratively discovering high-risk edge cases.
• Why Robots Still Struggle With Simple Tasks (And What Might Finally Change That) | Karol Hausman, Co-Founder & CEO of Physical Intelligence — Combining internet-scale LLM knowledge with robot motion data enables general physical intelligence for complex tasks.
• Robbyant Open Sources LingBot World: a Real Time World Model for Interactive Simulation and Embodied AI — LingBot World enables long-horizon, interactive video simulations for embodied AI using a 28B parameter diffusion transformer.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by ΑΙhub.