Saket Tiwari

Saket Tiwari

Researcher in reinforcement learning and robotic control

Posts by Tags

continual learning

Preserving Plasticity in Continual Robot Learning

12 minute read

Published:

Swapping AdamW for a Stiefel-constrained optimizer keeps pixel-based robot learners plastic for 100M frames, and a Gram-preserving variant improves continual fine-tuning of a VLA on a real SO-101 arm from 37% to 62% success.

optimization

Preserving Plasticity in Continual Robot Learning

12 minute read

Published:

Swapping AdamW for a Stiefel-constrained optimizer keeps pixel-based robot learners plastic for 100M frames, and a Gram-preserving variant improves continual fine-tuning of a VLA on a real SO-101 arm from 37% to 62% success.

plasticity

Preserving Plasticity in Continual Robot Learning

12 minute read

Published:

Swapping AdamW for a Stiefel-constrained optimizer keeps pixel-based robot learners plastic for 100M frames, and a Gram-preserving variant improves continual fine-tuning of a VLA on a real SO-101 arm from 37% to 62% success.

reinforcement learning

Preserving Plasticity in Continual Robot Learning

12 minute read

Published:

Swapping AdamW for a Stiefel-constrained optimizer keeps pixel-based robot learners plastic for 100M frames, and a Gram-preserving variant improves continual fine-tuning of a VLA on a real SO-101 arm from 37% to 62% success.

robotics

Preserving Plasticity in Continual Robot Learning

12 minute read

Published:

Swapping AdamW for a Stiefel-constrained optimizer keeps pixel-based robot learners plastic for 100M frames, and a Gram-preserving variant improves continual fine-tuning of a VLA on a real SO-101 arm from 37% to 62% success.