Research Intern: Test-Time Adaptation For Embodied Agents

Honda Research Institute USA; San Jose (CA); On-site; USD 33,000 - 50,000; Full time

Job summary

Honda Research Institute USA invites a Research Intern to explore test-time adaptation for embodied agents. Based in San Jose, CA, the intern will study when to revise representations or gather more information, using deployment evidence to guide updates.

The role emphasizes multimodal models, interaction, and continual learning in embodied AI The position targets Ph.D. or MS students in CS/ML with experience in PyTorch and relevant learning paradigms, offering a 3-month internship starting Jan

Qualifications

• Currently enrolled as a Ph.D. or MS student in Computer Science, Machine Learning, or a related field.
• Research experience in embodied learning, imitation learning, reinforcement learning, meta learning, transfer learning, lifelong learning.
• Experience with open-source ML/AI frameworks such as PyTorch.

Responsibilities

• Design, implement, and evaluate adaptive embodied AI agents that operate under uncertainty at test time.
• Use deployment evidence to revise an agent's task representation, state/world model, policy, or memory.
• Develop multimodal models integrating vision, language, and action for embodied/interactive settings.
• Build or extend experiments in embodied-AI and robot-learning environments.
• Compare adaptation mechanisms like in-context updates, retrieval, memory revision, and fast weights.
• Run controlled experiments and iterate on model and system design based on results.
• Document methodologies and findings in technical reports and papers.
• Collaborate with cross-disciplinary team members and participate in design reviews.

Skills

Embodied learning

Imitation learning

Reinforcement learning

Meta learning

Transfer learning

Lifelong learning

PyTorch

Education

Ph.D. or MS student in Computer Science or Machine Learning

Tools

PyTorch

Job description

Research Intern: Test-Time Adaptation For Embodied Agents - Honda Research Institute USA

Job Number: P25INT-59

Honda Research Institute USA (HRI-US) is seeking a highly motivated intern to investigate test-time adaptation for embodied AI. This project focuses on how evidence gathered during deployment can determine what an agent should revise, when it should gather more information or preserve its current model, and how useful corrections can persist across future interactions. The research will emphasize interaction-conditioned adaptation: learning from observations, action outcomes, failures, corrections, and human feedback collected while an agent operates. Potential directions include revising task or procedural representations, state and world models, policies, and memory through in-context adaptation, retrieval, parameter updates, fast weights, or other adaptation mechanisms. Embodied and robot-learning systems are a motivating application, including agents that learn procedures from demonstrations and refine them through experience. Experiments may use embodied-AI or robot-learning simulators, multimodal foundation models, vision-language-action policies, or related interactive-agent frameworks. This position is well suited to a student interested in test-time adaptation, continual learning, robot learning, multimodal models, or adaptive agent systems.

San Jose, CA

Key Responsibilities

• Design, implement, and evaluate adaptive embodied AI agents that operate under uncertainty and adjust behavior at test time based on context, feedback, or human interaction.
• Design and implement methods that use deployment evidence to revise an agent's task representation, state or world model, policy, memory, or another appropriate component in its decision-making apparatus.
• Develop and experiment with multimodal models and agents, integrating vision, language, and action for embodied or interactive settings.
• Build or extend experimental tasks in embodied-AI, robot-learning, or related interactive simulation environments
• Compare relevant adaptation mechanisms, such as in-context updates, retrieval, explicit memory revision, parameter-efficient optimization, fast weights, or world- and action-model adaptation.
• Run controlled experiments, analyze results, and iterate on model and system design based on empirical findings.
• Document methodologies and findings through technical reports, presentations, and research papers.
• Collaborate with cross-disciplinary team members and participate in research discussions and design reviews.

Minimum Qualifications

• Currently enrolled as a Ph.D. or M.S. student in Computer Science, Machine Learning, or a related field at a reputed university.
• Research experience in projects related to embodied learning, imitation learning, reinforcement learning, meta learning, transfer learning, lifelong learning.
• Experience in open-source ML/AI frameworks (PyTorch, etc.)

Bonus Qualifications

• Experience with embodied simulation platforms, video datasets, egocentric observations, or other robot policy learning experience.
• Familiarity with post-training multimodal AI models.
• Strong publication records in topics related to multi-agent system in robotics or AI conferences (RSS, CORL, NeurIPS, AAAI, AAMAS, ICLR).

Years of Work Experience Required 0

Desired Start Date 1/11/2027

Internship Duration 3 Months

Position Keywords Robot Learning, Embodied AI, Multimodal Foundation Models