Bilevel Optimization for Planning Through Contact: A Semidirect Method

Benoit, Landry; Lorenzetti, Joseph; Manchester, Zachary; Pavone, Marco

doi:10.1007/978-3-030-95459-8_48

Cited by 13 publications

(13 citation statements)

References 7 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Remark 5: The robot, controlled by the proposed jumping policy, shows the ability to not rely on the pre-defined contact plan and can break the contact after it lands and make contact again when it needs to utilize impacts to stabilize itself. Such a capability is similar to contact implicit trajectory optimization [3], [27]- [30]. While such optimization schemes still need to be computed offline for legged robots, our work achieves this online.…”

Section: Diverse and Robust Maneuvers By Multi-task Policymentioning

confidence: 98%

“…1) Model-based optimal control for legged jumping: Prior model-based methods for legged jumping control usually build up a layered optimization scheme, which includes offline trajectory optimization with detailed models of the robot's dynamics and ground contacts [15], [19]- [21], and online controllers that leverage simplified models of the robot's dynamics [6], [22]- [24]. In order to optimize trajectories for jumping, which needs to switch among modes with different underlying dynamics, there are two commonly employed solutions: relying on human-tuned predefined contact sequences [5], [12], [14], [25], [26], which is not scalable to different jump distances and/or directions, or leveraging contact-implicit optimization [3], [27]- [30] which plans through contacts to avoid breaking the trajectory or using computationally expensive mixed-integer programming [20], [31], [32]. However, due to the computational challenges of optimization, both of the above-mentioned methods are still limited to offline computation for legged robots.…”

Section: Related Workmentioning

confidence: 99%

See 1 more Smart Citation

Robust and Versatile Bipedal Jumping Control through Multi-Task Reinforcement Learning

Li¹,

Peng²,

Abbeel³

et al. 2023

Preprint

View full text Add to dashboard Cite

0°-55°( a) (b) (c) Flight Phase Fig. 1: Representative dynamic jumping maneuvers performed by a bipedal robot Cassie using the proposed multi-task control policies. From left to right: (a) the robot jumps over 1.4 m and lands at the given target; (b) the robot jumps to a target that is 0.88 m in front of the robot and 0.44 m above the ground, and (c) the robot jumps in place while turning 55 • with a command to turn 60 • in place. The policies are trained in simulation and deployed on the hardware without further tuning. Video is at: https://youtu.be/aAPSZ2QFB-E.

show abstract

Section: Diverse and Robust Maneuvers By Multi-task Policymentioning

confidence: 98%

Section: Related Workmentioning

confidence: 99%

Robust and Versatile Bipedal Jumping Control through Multi-Task Reinforcement Learning

Li¹,

Peng²,

Abbeel³

et al. 2023

Preprint

View full text Add to dashboard Cite

show abstract

“…Implicit models of contact using complementarity constraints [14], [15] are widely used in planning through contact for their ability to stably take bigger time steps [1], [38]. However, unlike penalty methods which permit explicit smoothing of forces [3], [19], [20], smoothing results of implicit optimization problems is less straightforward; randomized smoothing is unique in that it provides the ability to do so without an explicit relaxation of constraints.…”

Section: A Contacts Defined By Complementarity Constraintsmentioning

confidence: 99%

Bundled Gradients through Contact via Randomized Smoothing

Terry¹,

Pang²,

Tedrake³

2021

Preprint

View full text Add to dashboard Cite

The empirical success of derivative-free methods in reinforcement learning for planning through contact seems at odds with the perceived fragility of classical gradient-based optimization methods in these domains. What is causing this gap, and how might we use the answer to improve gradient-based methods? We believe a stochastic formulation of dynamics is one crucial ingredient. We use tools from randomized smoothing to analyze sampling-based approximations of the gradient, and formalize such approximations through the gradient bundle. We show that using the gradient bundle in lieu of the gradient mitigates fast-changing gradients of non-smooth contact dynamics modeled by the implicit time-stepping, or the penalty method. Finally, we apply the gradient bundle to optimal control using iLQR, introducing a novel algorithm which improves convergence over using exact gradients. Combining our algorithm with a convex implicit time-stepping formulation of contact, we show that we can tractably tackle planningthrough-contact problems in manipulation.

show abstract

“…These methods, termed CITO, simultaneously plan state, control input, and contact force trajectories without needing a pre-specified contact mode schedule. They handle the hybrid dynamics of contact with either complementarity constraints [2]- [6] or with soft constraints implemented as a penalty term in the cost function [7], [8]. In this work, we follow the formulation introduced in [2].…”

Section: A Contact-implicit Trajectory Optimizationmentioning

confidence: 99%

“…Since the introduction of CITO methods, many works have applied them to whole-body dynamic motion planning [9], [10], quadruped locomotion [6], and single-leg jumping [11]. Of these works, [10] and [11] use hierarchical planning schemes, first planning a trajectory with a simplified robot model and then using this to warm-start the full trajectory optimization.…”

Section: A Contact-implicit Trajectory Optimizationmentioning

confidence: 99%

TrajectoTree: Trajectory Optimization Meets Tree Search for Planning Multi-contact Dexterous Manipulation

Chen¹,

Culbertson²,

Lepert³

et al. 2021

Preprint

View full text Add to dashboard Cite

Dexterous manipulation tasks often require contact switching, where fingers make and break contact with the object. We propose a method that plans trajectories for dexterous manipulation tasks involving contact switching using contact-implicit trajectory optimization (CITO) augmented with a high-level discrete contact sequence planner. We first use the high-level planner to find a sequence of finger contact switches given a desired object trajectory. With this contact sequence plan, we impose additional constraints in the CITO problem. We show that our method finds trajectories approximately 7 times faster than a general CITO baseline for a four-finger planar manipulation scenario. Furthermore, when executing the planned trajectories in a full dynamics simulator, we are able to more closely track the object pose trajectories planned by our method than those planned by the baselines.

show abstract

Bilevel Optimization for Planning Through Contact: A Semidirect Method

Cited by 13 publications

References 7 publications

Robust and Versatile Bipedal Jumping Control through Multi-Task Reinforcement Learning

Robust and Versatile Bipedal Jumping Control through Multi-Task Reinforcement Learning

Bundled Gradients through Contact via Randomized Smoothing

TrajectoTree: Trajectory Optimization Meets Tree Search for Planning Multi-contact Dexterous Manipulation

Contact Info

Product

Resources

About