I am a B.Tech Computer Science & Biosciences student at Manipal University Jaipur, working on ML evaluation, vision-language models, and robotic manipulation.
My current research focus is on testing where frontier and open-weight models fail when they are pushed from language-level reasoning into grounded physical tasks. I am especially interested in VLM affordance prediction, robotic grasping, egocentric video understanding, and multilingual LLM behavior.
- Vision-language model evaluation for robotics and physical affordances
- Robotic manipulation, grasping, and part grounding
- Multilingual LLM evaluation, especially Hindi and Indian-language settings
- Egocentric video, IMU signals, and activity boundary detection
- Reliable benchmarks with simple baselines, reproducible code, and failure-mode analysis
- SmolWAM on RoboCasa: Moving a 0.88B world-action model from LIBERO to RoboCasa kitchens. Zero-shot it failed 30 of 30 episodes; after behavior-cloning fine-tuning on 3,108 demos it succeeds in 9 of 20 episodes in held-out kitchens. Includes the adapter, training recipe, results and videos.
- VLM Force Prediction: Zero-shot benchmark testing whether a VLM can infer robotic grasping force from images; includes physics-based ground truth, fragile-object safety analysis, and result scripts.
- Phase-Conditioned Trajectory: Research code for phase-conditioned Kalman filtering of hand trajectories and next-phase prediction experiments.
- WEAR IMU Segmentation: IMU vs. I3D vs. fusion experiments for subtask boundary detection using 18-subject LOSO cross-validation.
- IODarkWatch: Open-source maritime domain awareness project for detecting suspicious vessel behavior and AIS gaps in the Indian Ocean region.
- GroundBench: A factorized, counterfactual benchmark for locating where VLMs fail at affordance prediction. Poster, NeurIPS 2026 RoboPAD Workshop. arXiv:2609.13308
- Extending Beacon to Hindi: Cross-lingual sycophancy evaluation in English, Hindi literal translation, and Hindi culturally adapted settings. arXiv:2602.00046
- Part Grounding, Not Action Knowledge: VLM affordance prediction experiments on articulated-object manipulation. Currently under arXiv review/hold.
Python, PyTorch, MuJoCo, robosuite, RoboCasa, scikit-learn, OpenCV, MediaPipe, MiDaS, FilterPy, NumPy, Pandas, Matplotlib, Hugging Face Transformers, LM Studio, Git, LaTeX.
- Email: ssattigeri65@gmail.com
- GitHub: S1rlvk
- arXiv: 2609.13308, 2602.00046