← Writing

Robotics Research summary · 2026 · 9 min read

The HiRo Lab reading list: robots that learn from people

IISc’s HiRo Lab works on the theory and the algorithms that let robots learn from, adapt to, and collaborate with people in messy real-world settings. Their research page reads like a map of that problem. Here is a summary of every paper on it, grouped by the five threads they pull on.

The common thread is a robot that does not get a clean specification of the world. It gets a human teacher who is inconsistent, a language instruction that is vague, a gripper it has never seen, or a contact task where a wrong move hurts someone. The work below is roughly evenly split between making robots more capable and making them provably safe while they do it.

Foundation models for robotics

  • Gripper-aware Vision Language Action Models

    ECCV’26

    Vision-language-action policies quietly assume one gripper, but a parallel jaw and a suction cup need different strategies, and the training data skews to parallel jaws. G-VLA introduces MiGA, 103,000 demonstrations across five gripper types, plus a multi-gripper tokenizer and adapter-based policy routing, and transfers better to new objects, tasks and grippers in both simulation and the real world.

  • OVITA: Open-Vocabulary Interpretable Trajectory Adaptations

    RA-L’25

    Lets a non-expert reshape a robot’s trajectory with a plain-language instruction instead of code. It chains several pretrained LLMs to write the adaptation as code, with a separate LLM explaining what changed, demonstrated on a KUKA IIWA arm, a Clearpath Jackal ground robot and a Crazyflie drone.

Interactive RL and imitation learning

  • Beyond the Teacher: Leveraging Mixed-Skill Demonstrations for Robust Imitation Learning

    ICRA’26

    Aims for expert-level skill from imperfect, mixed-skill demonstrations rather than a large clean set. A two-stage pipeline scores and selects demonstrations, extracts a canonical motion with periodic Dynamic Movement Primitives, then refines it with an LSTM, reaching expert-like performance from as few as four demonstrations on wiping, weaving and pick-and-place.

  • PACER: Progress-Aligned Curation for Error-Resilient Imitation Learning

    CoRLW’25

    Targets imitation from sparse, noisy demonstrations where corrupted segments and misaligned timing wreck the policy. It learns a state-dependent task phase to align demonstrations in time, then uses a leave-one-out consensus to detect and down-weight corrupted segments, collapsing to a simple weighted behaviour-cloning step.

  • RISE: Robust Imitation through Stochastic Encodings

    IROSW’25

    Behaviour-cloned policies fail when the environment is measured noisily or gets disturbed at deployment. RISE encodes parameters such as obstacle state, orientation and velocity into a smooth variational latent space, so the imitation policy generalizes better at test time.

  • Impedance Primitive-Augmented Hierarchical Reinforcement Learning for Sequential Tasks

    ICRA’25

    Gives a hierarchical RL agent behaviour primitives that execute under variable-stiffness control, with an adaptive stiffness controller and affordance coupling. It learns faster and succeeds more often on contact-rich sequential tasks: block lifting, door opening, pushing and surface cleaning.

  • Sequentially Teaching Sequential Tasks (ST)²

    RA-M’26

    Long-horizon tasks are hard to demonstrate because deviations accumulate and human teachers fatigue. (ST)² lets a person set keypoints and teach incrementally; in a 16-participant study on a retail restocking task, the sequential approach beat teaching the whole task at once for most users, 10 to 5 with one tie.

Safe and compliant human-robot interaction

  • Safe and Optimal Variable Impedance Control via Certified Reinforcement Learning

    ICRA’26

    Letting RL tune impedance gains risks instability and unsafe exploration as those gains change over time. Certified Gaussian Manifold Sampling reframes exploration as drawing from a manifold of provably stable gain schedules, giving stable rollouts and bounded tracking error under model uncertainty, in simulation and on a real robot.

  • SafeDMPs: Integrating Formal Safety with DMPs for Adaptive HRI

    ICRA’26

    DMPs generalize from a single demonstration but carry no safety guarantee; control barrier functions are provably safe but lean on expensive real-time optimization. SafeDMPs adds Spatio-Temporal Tubes for a closed-form policy that dodges static and dynamic obstacles and runs orders of magnitude faster on a 7-DOF manipulator.

  • Safe Flow Q-Learning

    RLC’26

    Offline safe RL: find a reward-maximizing policy from a fixed dataset under strict safety constraints. SafeFQL pairs a Hamilton-Jacobi reachability safety value with behaviour cloning to train a flow policy, then distills it into a one-step actor, matching prior methods with fewer constraint violations and far lower inference latency than diffusion baselines.

  • V-OCBF: Value-Guided Offline Control Barrier Functions

    TMLR’26

    Learns a safety filter purely from fixed offline data, since expectation-based safe-RL constraints do not guarantee per-trajectory safety and classical barrier functions need hand-designed barriers and known dynamics. V-OCBF learns the dynamics and a neural barrier with a reachability Bellman target, then runs it as a real-time QP filter that cuts violations while preserving task performance.

Optimization and optimal control

3D vision, SLAM and world models

  • AffordMatcher: Affordance Learning in 3D Scenes from Visual Signifiers

    CVPR’26

    Finds the interactive regions in cluttered 3D indoor scenes, including the small but crucial parts like knobs and handles that object-level methods miss. It introduces AffordBridge, 291,637 interaction annotations over 685 scenes, and matches image to point-cloud keypoints, reaching 53.4 mAP against a 45.6 prior best at 20.7M parameters and 112.5 ms inference.

  • Beyond Visual Grasping: Benchmarking Complex Grasping from Detection to Execution

    IROS’26

    Grasping benchmarks usually test pose detection in isolation and skip the reasoning and reliable execution a real task needs. GCA-Bench spans 102 tasks across singulated, cluttered, constrained and semantic scenarios with 2,800 trajectories, and finds even recent foundation models and specialist grasping methods stay below 70% success on the hard cases.

Read together, the list is a bet that the frontier in robot learning is no longer raw capability but the interface with the world outside the lab: unreliable teachers, ambiguous language, unfamiliar hardware, and the safety guarantees you need before a robot shares space with a person. Full details and papers are on the HiRo Lab research page.