Agentic reasoning models trained with multimodal reinforcement learning (MMRL) have become increasingly capable, yet they are almost universally optimized using sparse, outcome-based rewards computed based on the final answers. Richer rewards computed from the reasoning tokens can improve learning significantly by providing more fine-grained guidance. However, it is challenging to compute more informative rewards in MMRL beyond those based on outcomes since different samples may require different scoring functions and teacher models may provide noisy reward signals too. In this paper, we introduce the Argos (Agentic Reward for Grounded & Objective Scoring), a principled reward agent to train multimodal reasoning models for agentic tasks. For each sample, Argos selects from a pool of teacher-model derived and rule-based scoring functions to simultaneously evaluate: (i) final response accuracy, (ii) spatiotemporal localization of referred entities and actions, and (iii) the quality of the reasoning process. We find that by leveraging our agentic verifier across both SFT data curation and RL training, our model achieves state-of-the-art results across multiple agentic tasks such as spatial reasoning, visual hallucination as well as robotics and embodied AI benchmarks. Critically, we demonstrate that just relying on SFT post-training on highly curated reasoning data is insufficient, as agents invariably collapse to ungrounded solutions during RL without our online verification. We also show that our agentic verifier can help to reduce reward-hacking in MMRL. Finally, we also provide a theoretical justification for the effectiveness of Argos through the concept of pareto-optimality.

Recent weeks have seen a number of reports of the death of self-driving cars, triggered mostly by the closure of Argo.AI, the self-driving startup funded by Ford and VW. That shutdown hasn't been the only bad news in a year of bad news for almost all of tech. The declarations that self-driving cars are a decade or more away have come from press like Bloomberg and Techcrunch, but perhaps more concerningly from insiders like Oxbotica and Luminar (though they claim to have always felt this way.) We've also seen all the public companies (except MobilEye) see very large stock declines this year, especially the SPAC-funded startups. Monday saw the "merger" of LIDAR pioneer Velodyne into Ouster.

June 2020: Volkswagen will invest USD 2.6 billion in capital and assets in Ford's self-driving unit ... [ ] Argo AI to market new-technology vehicles in the United States and Europe. Argo.ai, an autonomous vehicle (AV) unicorn, announced on October 26, 2022, that it would terminate operations across its locations in Pittsburgh, San Francisco, New Jersey and Munich. It impacts roughly 2100 employees and, more importantly, puts into focus the feasibility and business model for autonomous transportation on public roads. The company was founded in 2013 by alums of the 2006 DARPA Grand Challenge, Google, and Uber AV initiatives and secured significant backing and funding from Ford in 2017 ($1B) and later from Volkswagen in 2020 ($2.5B). Both investments occurred under the auspices of earlier CEOs, who have since been replaced (Ford CEO Jim Hacket in 2020 and Volkswagen CEO Herbert Deiss in 2022). It is unsurprising that new management saw things differently and did not want to inherit their predecessor's strategic bets.

DETROIT (Reuters) - The road map to fully self-driving vehicles is being rewritten once again, this time by Ford Motor Co and Volkswagen AG. When the two automakers joined forces in July 2019 to share control of self-driving startup Argo AI, it shook up the landscape among other key players. Wednesday's announcement that Pittsburgh-based Argo is being shuttered and some of its employees moving to Ford and VW underscores the growing realization that automated vehicles may be even further away from mass deployment than industry executives predicted back in 2019. "It's become very clear that profitable, fully autonomous vehicles at scale are still a long way off," Ford CFO John Lawler said on Wednesday. As Ford, General Motors Co and other companies began to realize they would need to step up investment over a longer period of time, "it was never clear what the financial returns were going to be" on automated vehicles, Evangelos Simoudis, an investor, author and corporate adviser, told Reuters on Wednesday.

Argo AI, the self-driving startup backed by Ford and Volkswagen, is shutting down, The Verge has learned. Employees were notified that an announcement would be made late in the day Wednesday. The company, which was founded by veterans of Google and Uber's self-driving car projects, has lost the financial support of Ford and VW, a source said. And according to TechCrunch, the company's resources will be absorbed by both automakers. Argo is estimated to have around 2,000 employees, though it did announce a round of layoffs earlier this year.

Self-driving car developer Argo AI suddenly announced that it was closing its doors this week. Some of its 1,800-odd employees, winnowed already by summer layoffs, are to be offered jobs to "work on automated technology with either Ford or Volkswagen," Catherine Johnsmeyer, an Argo spokesperson, said in a statement. The two auto giants had sunk some $3.6 billion into Argo and owned most of it. Now, they had decided to pull the plug. The end of Argo is just the latest sign that the global effort to get cars to drive themselves is in trouble--or at least more complex than once thought.

Canva uses a variety of machine learning (ML) models, such as recommender systems, information retrieval, attribution models, and natural language processing for various applications. A typical problem is the amount of time and engineering effort in choosing a set of optimal hyperparameters and configurations used to optimize a learning algorithm's performance. Hyperparameters are parameters set before a model's learning procedure begins. Hyperparameters, such as the learning rate and batch sizes, control the learning process and affect the predictive performance. Some hyperparameters might also have a significant impact on model size, inference throughput, latency, or other considerations. The number of hyperparameters in a model and their characteristics form a search space of possible combinations to optimize.

Over the last several decades, genetic research has seen incredible advances in gene sequencing technologies. In 2004, scientists completed the Human Genome Project, an ambitious project to sequence the human genome, which cost $3 billion and took 10 years. Now, a person can get their genome sequenced for less than $1,000 and within about 24 hours. Scientists capitalized on these advances by sequencing everything from the elusive giant squid to the Ethiopian eggplant. With this technology came promises of miraculous breakthroughs: all diseases would be cured and world hunger would be a thing of the past.

Ford-backed autonomous vehicle company Argo AI recently started driverless vehicle operations in both Miami, Florida, and Austin, Texas, during daylight hours. "Argo is first to go driverless in two major American cities, safely operating amongst heavy traffic, pedestrians and bicyclists in the busiest of neighborhoods," said Bryan Salesky, Founder and CEO of Argo AI. "From day one, we set out to tackle the hardest miles to drive -- in multiple cities -- because that's where the density of customer demand is, and where our autonomy platform is developing the intelligence required to scale it into a sustainable business." Here's a video Argo made about their new operations (article continues below): It took the company five years to get to the point where they could trust their vehicles to operate in busy traffic in these cities (two of eight cities they're now developing autonomous vehicles in). Driverless operations include multiple customer-facing operations, including work with several different commercial partners. For example, Argo has been working with Lyft to find riders in test vehicles since last year instead of offering their own rideshare platform and software.

Argo tests in multiple cities to ensure its SDS is exposed to a wide range of driving regulations, enabling it to operate appropriately and consistently with local rules, which often vary from place to place. Consider, for example, how a vehicle should behave when turning right if there is a bike lane. In California, a car may occupy the bike lane to turn right on red, but in Pennsylvania, the same right turn requires the car to stay in the vehicle lane. Argo's powerful prediction system can incorporate a database of driving styles from which to match data, anticipate likely actions, make appropriate decisions, and avoid extreme situations in order to achieve "naturalistic driving." The SDS can even handle the (in)famous "Pittsburgh left," an unwritten rule in Argo's home city which calls for oncoming traffic to give up the right-of-way and politely let left-turning vehicles turn against a green.