PhD candidate, McGill University & Mila · Montreal, Canada
Email CV Google Scholar GitHub LinkedIn
I am a final-year PhD candidate in Computer Science at McGill University and Mila (GPA 4.00), advised by Glen Berseth and Hsiu-Chin Lin.
My research is on post-training: teaching a pretrained model a new skill or a new setting without losing what it already knows. I study it on robot policies, where forgetting and unsafe exploration have physical costs. I began with model predictive control for bipeds, then safe reinforcement learning and on-robot fine-tuning for quadrupeds. More recently, I have turned to post-training methods from large language models, such as low-rank adaptation, KL-regularized fine-tuning and GRPO, and my current work is on policies up to a 7B vision-language-action model. The same questions drive LLM post-training: what to update, how far to move from the base model, and how to add a capability without losing the others.
My work spans humanoid whole-body tracking and loco-manipulation on the Unitree G1, and safe fine-tuning on a physical Unitree Go2 as a visiting researcher at the Technical University of Munich, supported by Majid Khadiv. Before my PhD, I completed a B.Sc. in Computer Science at Amirkabir University of Technology and worked on model predictive control for torque-controlled bipeds at the Max Planck Institute for Intelligent Systems, in the group of Ludovic Righetti (NYU).
McGill undergraduates interested in a robot-learning project are welcome to email me with a short note on their interests.
I am open to research internships in 2027.
In preparation
How does the choice of fine-tuning method trade off keeping a pretrained policy's behaviour against adapting to a new objective? The study compares LoRA, KL-anchored LoRA, a bounded residual and full fine-tuning across humanoid locomotion and loco-manipulation, NVIDIA's SONIC foundation motion tracker, and OpenVLA-OFT, a 7B vision-language-action model post-trained with GRPO on LIBERO.
Transactions on Machine Learning Research (TMLR), 2026 (accepted)
Interleaving a learned policy with a recovery controller biases every on-policy update, and the importance-sampling correction is undefined whenever that controller is deterministic. SafeExplorer is a drop-in PPO modification that remains unbiased in that case. In simulation (HalfCheetah, Ant, Unitree Go1; ten seeds) it had 21 to 255 times fewer training-time falls than PPO while matching its final reward.
@article{daneshmand2026safeexplorer,
title = {SafeExplorer: An Unbiased Mixed-Policy Gradient for Reinforcement Learning with Recovery Interventions},
author = {Daneshmand, Elham and Khadiv, Majid and Berseth, Glen and Lin, Hsiu-Chin},
journal = {Transactions on Machine Learning Research},
issn = {2835-8856},
year = {2026},
url = {https://openreview.net/forum?id=ppOrbSsMM0}
}Under review, IEEE-RAS Humanoids 2026 · Workshop poster, RLC 2026
Fine-tunes simulation-trained locomotion policies safely by combining low-rank adaptation (LoRA) with a recovery policy that takes over near unsafe states. In sim-to-sim transfer from Isaac Lab to MuJoCo on jump and trot, it cut fine-tuning time by 46.5% on average relative to full fine-tuning, with near-zero safety violations, and it was validated by fine-tuning a jump policy on a physical Unitree Go2.
@misc{daneshmand2026slowrl,
title = {SLowRL: Safe Low-Rank Adaptation Reinforcement Learning for Locomotion},
author = {Daneshmand, Elham and Omar, Shafeef and Berseth, Glen and Khadiv, Majid and Lin, Hsiu-Chin},
year = {2026},
eprint = {2603.17092},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2603.17092}
}
IEEE Robotics and Automation Letters (RA-L), 2021 · presented at ICRA 2021
Robot video featured in IEEE Spectrum Video Friday, October 2020
Two-level model predictive control: the upper level chooses swing-foot landing location and timing from centre-of-mass feedback, and the lower level generates dynamically consistent swing-foot trajectories under joint-torque and friction-cone limits. Robust walking under disturbances on the real biped Bolt.
@article{daneshmand2021variable,
title = {Variable Horizon {MPC} With Swing Foot Dynamics for Bipedal Walking Control},
author = {Daneshmand, Elham and Khadiv, Majid and Grimminger, Felix and Righetti, Ludovic},
journal = {IEEE Robotics and Automation Letters},
volume = {6},
number = {2},
pages = {2349--2356},
year = {2021},
doi = {10.1109/LRA.2021.3061381}
}
International Conference on Advanced Robotics (ICAR), 2021
Generates walking and running gaits by relaxing the fixed centre-of-mass height of the linear inverted pendulum with virtual constraints in centroidal space; a quadratic program adapts step location and timing under disturbances. Validated in simulation on the biped Bolt.
@inproceedings{boroujeni2021unified,
title = {A Unified Framework for Walking and Running of Bipedal Robots},
author = {Ghoddousi Boroujeni, Mahrokh and Daneshmand, Elham and Righetti, Ludovic and Khadiv, Majid},
booktitle = {2021 20th International Conference on Advanced Robotics (ICAR)},
pages = {396--403},
year = {2021},
doi = {10.1109/ICAR53236.2021.9659392}
}
NeurIPS 2026 Workshop on AI-Native Academia (accepted)
My reviewing agent placed 5th in the ICML 2026 Agent Review Competition, an internal competition held at Mila whose leaderboard ranked agents by how well their verdicts tracked the real ICML 2026 accept and reject decisions. The five top-ranked participants helped test the platform and joined its paper as co-authors.
Independent project, 2026 (in progress)
A hierarchical agent for the open-ended survival game Crafter: a high-level policy distilled from a language-model teacher chooses goals, and low-level skills learned from human play are fine-tuned with reinforcement learning.
ICML 2026 Agent Review Competition, Koala Science · 5th place
Language-model agents that review research papers; the competition leaderboard scored how well each agent's verdicts tracked the real ICML 2026 accept and reject decisions. The platform behind the competition is described in the Koala Science paper above.
Course project with Liam Paull, 2024
Using large language models to drive a Duckiebot from camera input, and to command quadruped locomotion in natural language.