Robots have the potential for a "ChatGPT moment" by learning through conversation and examples to perform complex tasks. Generalist robot models are increasing the credibility of this prospect by demonstrating versatility. The challenge remaining is in developing the learning, control, and economic frameworks for robots to operate in new environments. Similar to ChatGPT, robotics needs a persuasive encounter with versatility beyond just language to achieve this. Ultimately, the focus shifts to enabling machines to translate language into physical action effectively.
Sponsors and organizations are crucial for making ROSCon Global an event. All sponsors will be on-site at ROSCon Global to engage with attendees through booths, talks, and demos. Platinum sponsors include AMD, NVIDIA, and Intel, among others. Gold sponsors include Canonical, Clearpath, and Intel, along with several others. Silver and Bronze sponsors include companies like AgileX, Autodesk, and Autoware. Startup Alley features sponsors such as 10Things and Botbot.
Skild AI's S1 robot foundation model learns new, long-horizon tasks from a single video demonstration using in-context learning, avoiding the need for retraining. This model interprets demonstrated intent and sequences to execute unfamiliar tasks like plant potting and assembly without explicit programming. The system showed a significant advantage over other AI systems in learning complex, multi-step manipulations. This approach reduces the need for extensive manual training, potentially making experience from deployments useful for future model development. The collaboration leverages NVIDIA's AI infrastructure, simulation, and robotics frameworks to advance adaptable robot intelligence.
The robotaxi market is projected to reach $400 billion by 2035, requiring massive computing for safe, scalable driverless fleets. NVIDIA provides an open platform spanning AI training, simulation, and in-vehicle computing necessary for this development. This involves a three-computer solution: training, simulation/validation, and in-vehicle systems. Robotaxi leaders like Uber, Waymo, and others are adopting this stack to develop intelligent systems. Specific NVIDIA technologies like DGX, Omniverse, and DRIVE Hyperion are being integrated across major automotive partners to achieve autonomous deployment.
Dexterous manipulation is hindered by robots' lack of tactile feedback, which is crucial for handling deformable objects and small items. Vision-language-action (VLA) models guide robots in many tasks but struggle with fine-grained control because they ignore tactile information. Researchers are creating specialized tactile datasets and models to incorporate this sensory information into robot learning. By using expert submodels for fast tactile control, progress has been made in complex manipulation tasks with higher success rates. Future breakthroughs depend on scaling up diverse tactile data and developing methods to effectively fuse visual and tactile information.
Hazardous robotic exploration requires balancing the acquisition of information against the risk of causing failure. The paper introduces a behavioral valuation framework to decide when reducing uncertainty is worth the associated risk. This is achieved by incorporating a risk-augmented Behavioral Information objective based on Prelec probability weighting. This valuation creates boundaries between high-information/high-risk and lower-information/lower-risk paths. Risk-aware objectives demonstrated that this framework can reduce hazard exposure and robot losses compared to standard baselines.
The paper develops a kinematic construction linking serial propagation, parallel-mechanism closure, and rigid-platform point fields within a dual screw framework. It uses Lie-algebraic Bell recurrences to organize derivatives of the exponential factor for fixed joint axes. This method yields equivalent representations of the serial twist jet and an arbitrary-order triangular active-passive closure recurrence for parallel mechanisms. Validation using specific mechanical systems confirms that the formulation accurately reconstructs platform jets and their affine fields. The entire formulation is purely kinematic and holds at regular configurations.
A reinforcement learning framework is used to enable a robotic hand to perform complex piano repertoire with human-level expressivity. A graph-based optimization strategy guides the robot to generate fingering patterns mimicking natural human movements. The control system incorporates a physics-inspired acoustic model to accurately reproduce dynamic variations in keypress velocity. Quantitative evaluations show the expressive control model surpasses baselines in both finger morphology similarity and dynamic velocity accuracy. The resulting performances are perceived as expressive and indistinguishable from human performances by listeners.
ReactHuman is a physics-grounded benchmark designed to evaluate human-like reactive decision-making in embodied multimodal large language models. It tests a model's ability to react to sudden physical hazards in simulated household scenarios, going beyond passive physics evaluation. The benchmark uses 1,000+ reproducible scenes with exact ground truth from 240 Hz rigid-body simulation, including physically contradictory objects. A five-metric suite scores reactions based on reasonableness, safety, and physical grounding. Results indicate that models frequently fail to achieve reactive safety, often prioritizing appearance over motion and lacking correct interception points.
Planning trajectories with kinematic equality constraints restricts motion to a submanifold requiring special handling. The authors propose using analytic inverse kinematics (IK) to parameterize feasible configurations, as analytic IK functions are necessary for gradient-based optimization. They develop a method to compute IK gradients by leveraging the inverse function theorem and introduce a domain extension for reachability constraints. This approach maintains a gradient signal outside the reachable workspace. Numerical experiments and a hardware demonstration validate the efficacy of this method for trajectory planning.
IMLE-VLA introduces a method for fast, single-step action generation in Vision-Language-Action policies. It replaces iterative multi-step action heads with a single-step conditional generator trained via conditional Implicit Maximum Likelihood Estimation (cIMLE). This approach avoids multi-step sampling, which mitigates inference bottlenecks and results in significantly higher action throughput. IMLE-VLA achieves state-of-the-art performance on the LIBERO benchmark and demonstrates improved real-world motion smoothness and faster task completion. The method maintains robustness under test-time perturbations, confirming its ability to preserve generalization.
AI-based vehicle testing is challenged by the gap between passing simulated tests and real-world performance. The research trained two steering networks in CARLA: one with clear conditions and another with adverse conditions like fog or night. Formal verification, using bound propagation on trained weights, was employed to assess steering drift between captured images. This method identified potential failures in the clear-trained policy without further simulation testing. The study concludes that formal verification is a valuable addition to simulation for automated driving verification.
LTLDiff is a framework designed for multi-agent robotic manipulation by combining Finite Linear Temporal Logic (LTLf) specification learning with diffusion policies. The framework learns task-specific LTLf formulas from natural language instructions using a large language model. An embedding of this learned logic guides both data generation and diffusion-based policy training. This guidance encourages trajectories to adhere to required temporal and coordination constraints. Experiments show that LTLDiff improves task success rates in multi-agent manipulation tasks.